{
"pdf_info": [
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
107,
80,
435,
116
],
"lines": [
{
"bbox": [
106,
79,
437,
99
],
"spans": [
{
"bbox": [
106,
79,
437,
99
],
"score": 1.0,
"content": "GRAPH LAYOUTS AND GRAPH CONTRASTIVE",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
102,
413,
118
],
"spans": [
{
"bbox": [
105,
102,
413,
118
],
"score": 1.0,
"content": "LEARNING VIA NEIGHBOUR EMBEDDINGS",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
136,
244,
157
],
"lines": [
{
"bbox": [
113,
136,
201,
147
],
"spans": [
{
"bbox": [
113,
136,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
112,
146,
245,
159
],
"spans": [
{
"bbox": [
112,
146,
245,
159
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
277,
186,
335,
200
],
"spans": [
{
"bbox": [
277,
186,
335,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
143,
210,
468,
375
],
"lines": [
{
"bbox": [
141,
210,
470,
223
],
"spans": [
{
"bbox": [
141,
210,
470,
223
],
"score": 1.0,
"content": "In node-level graph representation learning, there are two distinct paradigms. One",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
221,
469,
234
],
"spans": [
{
"bbox": [
141,
221,
469,
234
],
"score": 1.0,
"content": "is known as graph layouts, where nodes are embedded into 2D space for visualiza-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
233,
469,
244
],
"spans": [
{
"bbox": [
141,
233,
469,
244
],
"score": 1.0,
"content": "tion purposes. Another is graph contrastive learning, where nodes are parametri-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
141,
243,
470,
255
],
"spans": [
{
"bbox": [
141,
243,
470,
255
],
"score": 1.0,
"content": "cally embedded into a high-dimensional vector space based on node features. In",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
254,
469,
266
],
"spans": [
{
"bbox": [
141,
254,
469,
266
],
"score": 1.0,
"content": "this work, we show that these two paradigms are intimately related, and that both",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
266,
469,
277
],
"spans": [
{
"bbox": [
141,
266,
469,
277
],
"score": 1.0,
"content": "can be successfully approached via neighbour embedding methods. First, we in-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
276,
470,
288
],
"spans": [
{
"bbox": [
141,
276,
200,
288
],
"score": 1.0,
"content": "troduce graph",
"type": "text"
},
{
"bbox": [
200,
277,
205,
286
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
205,
276,
470,
288
],
"score": 1.0,
"content": "-SNE for two-dimensional graph drawing, and show that the result-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
141,
287,
470,
300
],
"spans": [
{
"bbox": [
141,
287,
470,
300
],
"score": 1.0,
"content": "ing layouts outperform all existing algorithms in terms of local structure preser-",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
298,
470,
311
],
"spans": [
{
"bbox": [
141,
298,
236,
311
],
"score": 1.0,
"content": "vation, as measured by",
"type": "text"
},
{
"bbox": [
236,
298,
258,
308
],
"score": 0.45,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
258,
298,
470,
311
],
"score": 1.0,
"content": "classification accuracy. Second, we introduce graph",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
310,
469,
320
],
"spans": [
{
"bbox": [
141,
310,
469,
320
],
"score": 1.0,
"content": "contrastive neighbor embedding (graph CNE), which uses a fully-connected neu-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
319,
469,
333
],
"spans": [
{
"bbox": [
141,
319,
469,
333
],
"score": 1.0,
"content": "ral network (MLP) to transform graph node features into an embedding space by",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
331,
470,
343
],
"spans": [
{
"bbox": [
141,
331,
470,
343
],
"score": 1.0,
"content": "optimizing the contrastive InfoNCE objective. We show that graph CNE, while",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
142,
343,
469,
354
],
"spans": [
{
"bbox": [
142,
343,
469,
354
],
"score": 1.0,
"content": "being conceptually simpler than most existing graph contrastive learning meth-",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
142,
353,
469,
365
],
"spans": [
{
"bbox": [
142,
353,
469,
365
],
"score": 1.0,
"content": "ods, produces competitive node representations and outperforms state-of-the-art",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
362,
393,
377
],
"spans": [
{
"bbox": [
141,
362,
393,
377
],
"score": 1.0,
"content": "MLP-based methods in terms of linear classification accuracy.",
"type": "text"
}
],
"index": 19
}
],
"index": 12
},
{
"type": "title",
"bbox": [
108,
394,
206,
406
],
"lines": [
{
"bbox": [
105,
393,
208,
410
],
"spans": [
{
"bbox": [
105,
393,
208,
410
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
418,
505,
496
],
"lines": [
{
"bbox": [
105,
418,
506,
432
],
"spans": [
{
"bbox": [
105,
418,
506,
432
],
"score": 1.0,
"content": "Many real-world datasets, ranging from molecule structure to citation networks come in form of",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
429,
505,
443
],
"spans": [
{
"bbox": [
105,
429,
380,
443
],
"score": 1.0,
"content": "graphs. As graphs are abstract objects consisting of a set of nodes",
"type": "text"
},
{
"bbox": [
380,
430,
389,
439
],
"score": 0.78,
"content": "\\nu",
"type": "inline_equation"
},
{
"bbox": [
390,
429,
468,
443
],
"score": 1.0,
"content": "and a set of edges",
"type": "text"
},
{
"bbox": [
468,
430,
475,
439
],
"score": 0.72,
"content": "\\mathcal { E }",
"type": "inline_equation"
},
{
"bbox": [
475,
429,
505,
443
],
"score": 1.0,
"content": ", graph",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
439,
505,
453
],
"spans": [
{
"bbox": [
105,
439,
398,
453
],
"score": 1.0,
"content": "representation learning, i.e. embedding graph nodes into a vector space",
"type": "text"
},
{
"bbox": [
398,
440,
411,
451
],
"score": 0.88,
"content": "\\mathbb { R } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
412,
439,
505,
453
],
"score": 1.0,
"content": ", is a popular approach",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
450,
505,
465
],
"spans": [
{
"bbox": [
105,
450,
505,
465
],
"score": 1.0,
"content": "in machine learning. Traditionally, a distinction is made between graph layout (or graph drawing)",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
461,
505,
476
],
"spans": [
{
"bbox": [
105,
461,
250,
476
],
"score": 1.0,
"content": "methods, which embed nodes into",
"type": "text"
},
{
"bbox": [
250,
462,
263,
472
],
"score": 0.86,
"content": "\\mathbb { R } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
263,
461,
505,
476
],
"score": 1.0,
"content": "for visualization purposes, and graph contrastive learning",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
473,
505,
486
],
"spans": [
{
"bbox": [
105,
473,
505,
486
],
"score": 1.0,
"content": "methods, which use higher-dimensional embeddings more suitable for downstream analysis, such",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
483,
227,
497
],
"spans": [
{
"bbox": [
105,
483,
227,
497
],
"score": 1.0,
"content": "as classification or clustering.",
"type": "text"
}
],
"index": 27
}
],
"index": 24
},
{
"type": "text",
"bbox": [
107,
501,
504,
524
],
"lines": [
{
"bbox": [
105,
501,
505,
514
],
"spans": [
{
"bbox": [
105,
501,
158,
514
],
"score": 1.0,
"content": "For a graph",
"type": "text"
},
{
"bbox": [
158,
501,
210,
513
],
"score": 0.93,
"content": "G = ( \\nu , \\mathcal { E } )",
"type": "inline_equation"
},
{
"bbox": [
210,
501,
505,
514
],
"score": 1.0,
"content": ", graph layout methods usually only take into account its structure and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
511,
505,
524
],
"spans": [
{
"bbox": [
106,
511,
505,
524
],
"score": 1.0,
"content": "obtain the layout by pulling together connected nodes. In contrast, graph contrastive learning (GCL)",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
},
{
"type": "image",
"bbox": [
119,
543,
498,
667
],
"blocks": [
{
"type": "image_body",
"bbox": [
119,
543,
498,
667
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
543,
498,
667
],
"spans": [
{
"bbox": [
119,
543,
498,
667
],
"score": 0.932,
"type": "image",
"image_path": "774b4b73abf66e7348ef56ad4f320e1e7f9e39a429d420e356684d04036a786b.jpg"
}
]
}
],
"index": 31,
"virtual_lines": [
{
"bbox": [
119,
543,
498,
584.3333333333334
],
"spans": [],
"index": 30
},
{
"bbox": [
119,
584.3333333333334,
498,
625.6666666666667
],
"spans": [],
"index": 31
},
{
"bbox": [
119,
625.6666666666667,
498,
667.0000000000001
],
"spans": [],
"index": 32
}
]
},
{
"type": "image_caption",
"bbox": [
106,
685,
505,
730
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
685,
506,
698
],
"spans": [
{
"bbox": [
105,
685,
506,
698
],
"score": 1.0,
"content": "Figure 1: 2D embeddings of the Amazon Computer (ACO) and Photo (APH) datasets, obtained",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
696,
506,
709
],
"spans": [
{
"bbox": [
105,
696,
173,
709
],
"score": 1.0,
"content": "using our graph",
"type": "text"
},
{
"bbox": [
173,
698,
178,
707
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
179,
696,
299,
709
],
"score": 1.0,
"content": "-SNE and graph CNE. Graph",
"type": "text"
},
{
"bbox": [
299,
698,
304,
707
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
304,
696,
506,
709
],
"score": 1.0,
"content": "-SNE is a graph layout method. Graph CNE is a",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
708,
506,
720
],
"spans": [
{
"bbox": [
105,
708,
506,
720
],
"score": 1.0,
"content": "graph contrastive learning method, mapping node features to 2D (or other dimensionality, e.g. 128D,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
718,
471,
731
],
"spans": [
{
"bbox": [
105,
718,
471,
731
],
"score": 1.0,
"content": "see Section 6) using a neural network. Embeddings were aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5
}
],
"index": 32.75
}
],
"page_idx": 0,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
302,
751,
309,
762
],
"spans": [
{
"bbox": [
302,
751,
309,
762
],
"score": 1.0,
"content": "1",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
107,
80,
435,
116
],
"lines": [
{
"bbox": [
106,
79,
437,
99
],
"spans": [
{
"bbox": [
106,
79,
437,
99
],
"score": 1.0,
"content": "GRAPH LAYOUTS AND GRAPH CONTRASTIVE",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
102,
413,
118
],
"spans": [
{
"bbox": [
105,
102,
413,
118
],
"score": 1.0,
"content": "LEARNING VIA NEIGHBOUR EMBEDDINGS",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "text",
"bbox": [
112,
136,
244,
157
],
"lines": [
{
"bbox": [
113,
136,
201,
147
],
"spans": [
{
"bbox": [
113,
136,
201,
147
],
"score": 1.0,
"content": "Anonymous authors",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
112,
146,
245,
159
],
"spans": [
{
"bbox": [
112,
146,
245,
159
],
"score": 1.0,
"content": "Paper under double-blind review",
"type": "text"
}
],
"index": 3
}
],
"index": 2.5,
"bbox_fs": [
112,
136,
245,
159
]
},
{
"type": "title",
"bbox": [
278,
186,
333,
199
],
"lines": [
{
"bbox": [
277,
186,
335,
200
],
"spans": [
{
"bbox": [
277,
186,
335,
200
],
"score": 1.0,
"content": "ABSTRACT",
"type": "text"
}
],
"index": 4
}
],
"index": 4
},
{
"type": "text",
"bbox": [
143,
210,
468,
375
],
"lines": [
{
"bbox": [
141,
210,
470,
223
],
"spans": [
{
"bbox": [
141,
210,
470,
223
],
"score": 1.0,
"content": "In node-level graph representation learning, there are two distinct paradigms. One",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
141,
221,
469,
234
],
"spans": [
{
"bbox": [
141,
221,
469,
234
],
"score": 1.0,
"content": "is known as graph layouts, where nodes are embedded into 2D space for visualiza-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
141,
233,
469,
244
],
"spans": [
{
"bbox": [
141,
233,
469,
244
],
"score": 1.0,
"content": "tion purposes. Another is graph contrastive learning, where nodes are parametri-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
141,
243,
470,
255
],
"spans": [
{
"bbox": [
141,
243,
470,
255
],
"score": 1.0,
"content": "cally embedded into a high-dimensional vector space based on node features. In",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
141,
254,
469,
266
],
"spans": [
{
"bbox": [
141,
254,
469,
266
],
"score": 1.0,
"content": "this work, we show that these two paradigms are intimately related, and that both",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
141,
266,
469,
277
],
"spans": [
{
"bbox": [
141,
266,
469,
277
],
"score": 1.0,
"content": "can be successfully approached via neighbour embedding methods. First, we in-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
141,
276,
470,
288
],
"spans": [
{
"bbox": [
141,
276,
200,
288
],
"score": 1.0,
"content": "troduce graph",
"type": "text"
},
{
"bbox": [
200,
277,
205,
286
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
205,
276,
470,
288
],
"score": 1.0,
"content": "-SNE for two-dimensional graph drawing, and show that the result-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
141,
287,
470,
300
],
"spans": [
{
"bbox": [
141,
287,
470,
300
],
"score": 1.0,
"content": "ing layouts outperform all existing algorithms in terms of local structure preser-",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
141,
298,
470,
311
],
"spans": [
{
"bbox": [
141,
298,
236,
311
],
"score": 1.0,
"content": "vation, as measured by",
"type": "text"
},
{
"bbox": [
236,
298,
258,
308
],
"score": 0.45,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
258,
298,
470,
311
],
"score": 1.0,
"content": "classification accuracy. Second, we introduce graph",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
141,
310,
469,
320
],
"spans": [
{
"bbox": [
141,
310,
469,
320
],
"score": 1.0,
"content": "contrastive neighbor embedding (graph CNE), which uses a fully-connected neu-",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
141,
319,
469,
333
],
"spans": [
{
"bbox": [
141,
319,
469,
333
],
"score": 1.0,
"content": "ral network (MLP) to transform graph node features into an embedding space by",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
141,
331,
470,
343
],
"spans": [
{
"bbox": [
141,
331,
470,
343
],
"score": 1.0,
"content": "optimizing the contrastive InfoNCE objective. We show that graph CNE, while",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
142,
343,
469,
354
],
"spans": [
{
"bbox": [
142,
343,
469,
354
],
"score": 1.0,
"content": "being conceptually simpler than most existing graph contrastive learning meth-",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
142,
353,
469,
365
],
"spans": [
{
"bbox": [
142,
353,
469,
365
],
"score": 1.0,
"content": "ods, produces competitive node representations and outperforms state-of-the-art",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
141,
362,
393,
377
],
"spans": [
{
"bbox": [
141,
362,
393,
377
],
"score": 1.0,
"content": "MLP-based methods in terms of linear classification accuracy.",
"type": "text"
}
],
"index": 19
}
],
"index": 12,
"bbox_fs": [
141,
210,
470,
377
]
},
{
"type": "title",
"bbox": [
108,
394,
206,
406
],
"lines": [
{
"bbox": [
105,
393,
208,
410
],
"spans": [
{
"bbox": [
105,
393,
208,
410
],
"score": 1.0,
"content": "1 INTRODUCTION",
"type": "text"
}
],
"index": 20
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
418,
505,
496
],
"lines": [
{
"bbox": [
105,
418,
506,
432
],
"spans": [
{
"bbox": [
105,
418,
506,
432
],
"score": 1.0,
"content": "Many real-world datasets, ranging from molecule structure to citation networks come in form of",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
429,
505,
443
],
"spans": [
{
"bbox": [
105,
429,
380,
443
],
"score": 1.0,
"content": "graphs. As graphs are abstract objects consisting of a set of nodes",
"type": "text"
},
{
"bbox": [
380,
430,
389,
439
],
"score": 0.78,
"content": "\\nu",
"type": "inline_equation"
},
{
"bbox": [
390,
429,
468,
443
],
"score": 1.0,
"content": "and a set of edges",
"type": "text"
},
{
"bbox": [
468,
430,
475,
439
],
"score": 0.72,
"content": "\\mathcal { E }",
"type": "inline_equation"
},
{
"bbox": [
475,
429,
505,
443
],
"score": 1.0,
"content": ", graph",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
439,
505,
453
],
"spans": [
{
"bbox": [
105,
439,
398,
453
],
"score": 1.0,
"content": "representation learning, i.e. embedding graph nodes into a vector space",
"type": "text"
},
{
"bbox": [
398,
440,
411,
451
],
"score": 0.88,
"content": "\\mathbb { R } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
412,
439,
505,
453
],
"score": 1.0,
"content": ", is a popular approach",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
450,
505,
465
],
"spans": [
{
"bbox": [
105,
450,
505,
465
],
"score": 1.0,
"content": "in machine learning. Traditionally, a distinction is made between graph layout (or graph drawing)",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
461,
505,
476
],
"spans": [
{
"bbox": [
105,
461,
250,
476
],
"score": 1.0,
"content": "methods, which embed nodes into",
"type": "text"
},
{
"bbox": [
250,
462,
263,
472
],
"score": 0.86,
"content": "\\mathbb { R } ^ { 2 }",
"type": "inline_equation"
},
{
"bbox": [
263,
461,
505,
476
],
"score": 1.0,
"content": "for visualization purposes, and graph contrastive learning",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
473,
505,
486
],
"spans": [
{
"bbox": [
105,
473,
505,
486
],
"score": 1.0,
"content": "methods, which use higher-dimensional embeddings more suitable for downstream analysis, such",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
483,
227,
497
],
"spans": [
{
"bbox": [
105,
483,
227,
497
],
"score": 1.0,
"content": "as classification or clustering.",
"type": "text"
}
],
"index": 27
}
],
"index": 24,
"bbox_fs": [
105,
418,
506,
497
]
},
{
"type": "text",
"bbox": [
107,
501,
504,
524
],
"lines": [
{
"bbox": [
105,
501,
505,
514
],
"spans": [
{
"bbox": [
105,
501,
158,
514
],
"score": 1.0,
"content": "For a graph",
"type": "text"
},
{
"bbox": [
158,
501,
210,
513
],
"score": 0.93,
"content": "G = ( \\nu , \\mathcal { E } )",
"type": "inline_equation"
},
{
"bbox": [
210,
501,
505,
514
],
"score": 1.0,
"content": ", graph layout methods usually only take into account its structure and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
511,
505,
524
],
"spans": [
{
"bbox": [
106,
511,
505,
524
],
"score": 1.0,
"content": "obtain the layout by pulling together connected nodes. In contrast, graph contrastive learning (GCL)",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
105,
501,
505,
524
]
},
{
"type": "image",
"bbox": [
119,
543,
498,
667
],
"blocks": [
{
"type": "image_body",
"bbox": [
119,
543,
498,
667
],
"group_id": 0,
"lines": [
{
"bbox": [
119,
543,
498,
667
],
"spans": [
{
"bbox": [
119,
543,
498,
667
],
"score": 0.932,
"type": "image",
"image_path": "774b4b73abf66e7348ef56ad4f320e1e7f9e39a429d420e356684d04036a786b.jpg"
}
]
}
],
"index": 31,
"virtual_lines": [
{
"bbox": [
119,
543,
498,
584.3333333333334
],
"spans": [],
"index": 30
},
{
"bbox": [
119,
584.3333333333334,
498,
625.6666666666667
],
"spans": [],
"index": 31
},
{
"bbox": [
119,
625.6666666666667,
498,
667.0000000000001
],
"spans": [],
"index": 32
}
]
},
{
"type": "image_caption",
"bbox": [
106,
685,
505,
730
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
685,
506,
698
],
"spans": [
{
"bbox": [
105,
685,
506,
698
],
"score": 1.0,
"content": "Figure 1: 2D embeddings of the Amazon Computer (ACO) and Photo (APH) datasets, obtained",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
696,
506,
709
],
"spans": [
{
"bbox": [
105,
696,
173,
709
],
"score": 1.0,
"content": "using our graph",
"type": "text"
},
{
"bbox": [
173,
698,
178,
707
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
179,
696,
299,
709
],
"score": 1.0,
"content": "-SNE and graph CNE. Graph",
"type": "text"
},
{
"bbox": [
299,
698,
304,
707
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
304,
696,
506,
709
],
"score": 1.0,
"content": "-SNE is a graph layout method. Graph CNE is a",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
708,
506,
720
],
"spans": [
{
"bbox": [
105,
708,
506,
720
],
"score": 1.0,
"content": "graph contrastive learning method, mapping node features to 2D (or other dimensionality, e.g. 128D,",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
718,
471,
731
],
"spans": [
{
"bbox": [
105,
718,
471,
731
],
"score": 1.0,
"content": "see Section 6) using a neural network. Embeddings were aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 36
}
],
"index": 34.5
}
],
"index": 32.75
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
137
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
252,
95
],
"score": 1.0,
"content": "methods typically use node features",
"type": "text"
},
{
"bbox": [
252,
83,
262,
93
],
"score": 0.3,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
263,
83,
292,
95
],
"score": 1.0,
"content": "of size",
"type": "text"
},
{
"bbox": [
292,
83,
321,
93
],
"score": 0.91,
"content": "n \\times D",
"type": "inline_equation"
},
{
"bbox": [
321,
83,
348,
95
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
349,
83,
382,
94
],
"score": 0.93,
"content": "n = | \\mathcal { V } |",
"type": "inline_equation"
},
{
"bbox": [
383,
83,
505,
95
],
"score": 1.0,
"content": "and employ a neural network,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
104,
90,
507,
109
],
"spans": [
{
"bbox": [
104,
90,
418,
109
],
"score": 1.0,
"content": "usually a graph convolutional network (GCN) (Kipf & Welling, 2017), for the",
"type": "text"
},
{
"bbox": [
418,
93,
464,
104
],
"score": 0.91,
"content": "\\mathbf { R } ^ { D } \\to \\mathbf { R } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
465,
90,
507,
109
],
"score": 1.0,
"content": "mapping.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
102,
506,
118
],
"spans": [
{
"bbox": [
105,
102,
506,
118
],
"score": 1.0,
"content": "GCL methods also pull connected nodes together, sometimes explicitly through their loss function,",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
114,
505,
129
],
"spans": [
{
"bbox": [
105,
114,
505,
129
],
"score": 1.0,
"content": "but also implicitly through the GCN architecture (Trivedi et al., 2022; Wang et al., 2023; Guo et al.,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
136,
140
],
"spans": [
{
"bbox": [
105,
124,
136,
140
],
"score": 1.0,
"content": "2023).",
"type": "text"
}
],
"index": 4
}
],
"index": 2
},
{
"type": "text",
"bbox": [
106,
143,
505,
221
],
"lines": [
{
"bbox": [
105,
142,
506,
156
],
"spans": [
{
"bbox": [
105,
142,
506,
156
],
"score": 1.0,
"content": "Recent work (Kruiger et al., 2017; Zhu et al., 2020a; Zhong et al., 2023; Bohm et al., 2022) pointed ¨",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
154,
505,
167
],
"spans": [
{
"bbox": [
105,
154,
476,
167
],
"score": 1.0,
"content": "out deep connections between graph layout and neighbor embedding algorithms such as",
"type": "text"
},
{
"bbox": [
477,
155,
482,
164
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
154,
505,
167
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
165,
504,
178
],
"spans": [
{
"bbox": [
106,
165,
504,
178
],
"score": 1.0,
"content": "(Van der Maaten & Hinton, 2008) or UMAP (McInnes et al., 2018), which are based on neigh-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
177,
504,
188
],
"spans": [
{
"bbox": [
106,
177,
504,
188
],
"score": 1.0,
"content": "borhood preservation. In parallel, another line of work explored connections between neighbor",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
200
],
"spans": [
{
"bbox": [
105,
187,
505,
200
],
"score": 1.0,
"content": "embeddings and contrastive learning (Damrich et al., 2022; Bohm et al., 2023; Hu et al., 2023). ¨",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
198,
505,
211
],
"spans": [
{
"bbox": [
105,
198,
505,
211
],
"score": 1.0,
"content": "This raises the question to what extent neighbor embedding and contrastive neighbor embedding",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
208,
403,
223
],
"spans": [
{
"bbox": [
106,
208,
403,
223
],
"score": 1.0,
"content": "algorithms (see Section 3) can be useful for graph representation learning.",
"type": "text"
}
],
"index": 11
}
],
"index": 8
},
{
"type": "text",
"bbox": [
107,
226,
505,
292
],
"lines": [
{
"bbox": [
105,
226,
505,
239
],
"spans": [
{
"bbox": [
105,
226,
477,
239
],
"score": 1.0,
"content": "In this work, we answer this question. We introduce a novel graph layout algorithm, graph",
"type": "text"
},
{
"bbox": [
477,
227,
482,
236
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
226,
505,
239
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
237,
505,
249
],
"spans": [
{
"bbox": [
105,
237,
505,
249
],
"score": 1.0,
"content": "(Figure 1), and show that it strongly outperforms existing methods. We also introduce a novel,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
247,
506,
261
],
"spans": [
{
"bbox": [
105,
247,
506,
261
],
"score": 1.0,
"content": "augmentation-free, GCL algorithm, graph CNE (Figure 1), based on the framework for contrastive",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
259,
505,
271
],
"spans": [
{
"bbox": [
106,
259,
505,
271
],
"score": 1.0,
"content": "neighbor embeddings, and show that it reaches competitive GCL performance without using GCNs.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
268,
506,
284
],
"spans": [
{
"bbox": [
105,
268,
506,
284
],
"score": 1.0,
"content": "Conceptually, we present a single coherent framework for node-level graph representation learning,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
281,
440,
294
],
"spans": [
{
"bbox": [
106,
281,
440,
294
],
"score": 1.0,
"content": "tying together graph layouts, graph contrastive learning, and neighbor embeddings.",
"type": "text"
}
],
"index": 17
}
],
"index": 14.5
},
{
"type": "title",
"bbox": [
108,
307,
211,
320
],
"lines": [
{
"bbox": [
104,
306,
213,
323
],
"spans": [
{
"bbox": [
104,
306,
213,
323
],
"score": 1.0,
"content": "2 RELATED WORK",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
107,
332,
505,
421
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Graph layouts Graph layout algorithms have traditionally been based on spring models, where",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
343,
505,
356
],
"spans": [
{
"bbox": [
105,
343,
332,
356
],
"score": 1.0,
"content": "every connected pair of nodes feels an attractive force",
"type": "text"
},
{
"bbox": [
333,
344,
345,
355
],
"score": 0.9,
"content": "F _ { a }",
"type": "inline_equation"
},
{
"bbox": [
345,
343,
505,
356
],
"score": 1.0,
"content": "and all pairs of nodes feel a repulsive",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
353,
504,
370
],
"spans": [
{
"bbox": [
104,
353,
129,
370
],
"score": 1.0,
"content": "force",
"type": "text"
},
{
"bbox": [
129,
355,
141,
366
],
"score": 0.86,
"content": "F _ { r }",
"type": "inline_equation"
},
{
"bbox": [
142,
353,
409,
370
],
"score": 1.0,
"content": "(force-directed graph layouts). Many algorithms can be written as",
"type": "text"
},
{
"bbox": [
410,
355,
447,
368
],
"score": 0.92,
"content": "F _ { a } = d _ { i j } ^ { a }",
"type": "inline_equation"
},
{
"bbox": [
448,
353,
466,
370
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
466,
355,
504,
368
],
"score": 0.92,
"content": "F _ { r } = d _ { i j } ^ { r }",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
105,
365,
505,
379
],
"spans": [
{
"bbox": [
105,
365,
198,
379
],
"score": 1.0,
"content": "(Noack, 2007), where",
"type": "text"
},
{
"bbox": [
198,
366,
212,
378
],
"score": 0.9,
"content": "d _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
212,
365,
505,
379
],
"score": 1.0,
"content": "is the embedding distance between nodes. For example, Fruchterman–",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
377,
505,
389
],
"spans": [
{
"bbox": [
105,
377,
296,
389
],
"score": 1.0,
"content": "Reingold algorithm, also known as FDP, uses",
"type": "text"
},
{
"bbox": [
297,
377,
363,
388
],
"score": 0.92,
"content": "a = 2 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
363,
377,
505,
389
],
"score": 1.0,
"content": "(Fruchterman & Reingold, 1991);",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
387,
506,
401
],
"spans": [
{
"bbox": [
105,
387,
177,
401
],
"score": 1.0,
"content": "ForceAtlas2 uses",
"type": "text"
},
{
"bbox": [
178,
388,
238,
399
],
"score": 0.92,
"content": "a = 1 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
238,
387,
380,
401
],
"score": 1.0,
"content": "(Jacomy et al., 2014); LinLog uses",
"type": "text"
},
{
"bbox": [
381,
388,
441,
399
],
"score": 0.93,
"content": "a = 0 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
441,
387,
506,
401
],
"score": 1.0,
"content": "(Noack, 2007).",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
398,
506,
411
],
"spans": [
{
"bbox": [
105,
398,
506,
411
],
"score": 1.0,
"content": "Efficient implementations can be based on Barnes–Hut approximation of the repulsive forces, as in",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
410,
490,
422
],
"spans": [
{
"bbox": [
106,
410,
490,
422
],
"score": 1.0,
"content": "SFDP (Hu, 2005). Relationship to neighbour embeddings was discussed by Bohm et al. (2022). ¨",
"type": "text"
}
],
"index": 26
}
],
"index": 22.5
},
{
"type": "text",
"bbox": [
107,
432,
505,
532
],
"lines": [
{
"bbox": [
106,
433,
505,
446
],
"spans": [
{
"bbox": [
106,
433,
225,
446
],
"score": 1.0,
"content": "Graph layouts inspired by",
"type": "text"
},
{
"bbox": [
226,
434,
231,
443
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
231,
433,
505,
446
],
"score": 1.0,
"content": "-SNE Several recent graph layout algorithms have been inspired",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
443,
506,
457
],
"spans": [
{
"bbox": [
105,
443,
450,
457
],
"score": 1.0,
"content": "by neighbor embeddings. tsNET (Kruiger et al., 2017) applied modified version of",
"type": "text"
},
{
"bbox": [
450,
444,
455,
454
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
455,
443,
506,
457
],
"score": 1.0,
"content": "-SNE to the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
104,
454,
506,
468
],
"spans": [
{
"bbox": [
104,
454,
506,
468
],
"score": 1.0,
"content": "pairwise shortest path distances between all nodes. DRGraph (Zhu et al., 2020a) made tsNET faster",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
465,
504,
479
],
"spans": [
{
"bbox": [
105,
465,
309,
479
],
"score": 1.0,
"content": "by using negative sampling (Mikolov et al., 2013).",
"type": "text"
},
{
"bbox": [
309,
466,
315,
476
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
315,
465,
491,
479
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023) suggested custom",
"type": "text"
},
{
"bbox": [
491,
466,
504,
477
],
"score": 0.88,
"content": "F _ { a }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
477,
505,
490
],
"spans": [
{
"bbox": [
105,
477,
123,
490
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
123,
477,
135,
488
],
"score": 0.88,
"content": "F _ { r }",
"type": "inline_equation"
},
{
"bbox": [
135,
477,
209,
490
],
"score": 1.0,
"content": "forces inspired by",
"type": "text"
},
{
"bbox": [
209,
478,
214,
487
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
214,
477,
505,
490
],
"score": 1.0,
"content": "-SNE and adopted interpolation-based approximation of Linderman et al.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
486,
505,
502
],
"spans": [
{
"bbox": [
105,
486,
284,
502
],
"score": 1.0,
"content": "(2019). Below we will show that our graph",
"type": "text"
},
{
"bbox": [
284,
488,
289,
497
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
289,
486,
443,
502
],
"score": 1.0,
"content": "-SNE outperforms both DRGraph and",
"type": "text"
},
{
"bbox": [
443,
488,
448,
498
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
448,
486,
505,
502
],
"score": 1.0,
"content": "-FDP. Finally,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
499,
505,
511
],
"spans": [
{
"bbox": [
106,
499,
346,
511
],
"score": 1.0,
"content": "Leow et al. (2019) also suggested an algorithm called ‘graph",
"type": "text"
},
{
"bbox": [
346,
500,
351,
509
],
"score": 0.57,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
351,
499,
505,
511
],
"score": 1.0,
"content": "-SNE’, that used a graph convolutional",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
508,
506,
524
],
"spans": [
{
"bbox": [
104,
508,
477,
524
],
"score": 1.0,
"content": "network (Kipf & Welling, 2017) to build a parametric mapping optimizing a combination of",
"type": "text"
},
{
"bbox": [
477,
510,
482,
520
],
"score": 0.73,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
508,
506,
524
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
520,
505,
533
],
"spans": [
{
"bbox": [
105,
520,
474,
533
],
"score": 1.0,
"content": "losses on node features and on shortest graph distances; it has almost no relation to our graph",
"type": "text"
},
{
"bbox": [
474,
522,
479,
530
],
"score": 0.65,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
480,
520,
505,
533
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 35
}
],
"index": 31
},
{
"type": "text",
"bbox": [
107,
543,
505,
676
],
"lines": [
{
"bbox": [
105,
543,
505,
557
],
"spans": [
{
"bbox": [
105,
543,
505,
557
],
"score": 1.0,
"content": "Node-level graph contrastive learning The basic principle behind contrastive learning is to learn",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
555,
505,
568
],
"spans": [
{
"bbox": [
105,
555,
505,
568
],
"score": 1.0,
"content": "data representation by contrasting pairs of observations that are similar to each other (positive pairs)",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
565,
505,
579
],
"spans": [
{
"bbox": [
105,
565,
505,
579
],
"score": 1.0,
"content": "with those that are dissimilar to each other (negative pairs). In computer vision, positive pairs are",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
576,
506,
591
],
"spans": [
{
"bbox": [
104,
576,
506,
591
],
"score": 1.0,
"content": "generated via data augmentation, e.g. in SimCLR (Chen et al., 2020). Graph contrastive learning",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
588,
505,
600
],
"spans": [
{
"bbox": [
105,
588,
505,
600
],
"score": 1.0,
"content": "(GCL) requires node features (as input to the network) and can be graph-level or node-level, de-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
599,
505,
612
],
"spans": [
{
"bbox": [
105,
599,
505,
612
],
"score": 1.0,
"content": "pending on whether representations are obtained for a set of graphs or for the set of nodes of a single",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
104,
609,
506,
623
],
"spans": [
{
"bbox": [
104,
609,
506,
623
],
"score": 1.0,
"content": "graph. Graph-level GCL is based on graph augmentations, such as node dropping or edge pertur-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
620,
506,
633
],
"spans": [
{
"bbox": [
105,
620,
506,
633
],
"score": 1.0,
"content": "bation, e.g. in GraphCL (You et al., 2020). Prominent examples of node-level GCL algorithms that",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
632,
505,
644
],
"spans": [
{
"bbox": [
105,
632,
505,
644
],
"score": 1.0,
"content": "are also based on graph augmentations include GRACE (Zhu et al., 2020b), GCA (Zhu et al., 2021),",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
641,
505,
655
],
"spans": [
{
"bbox": [
105,
641,
505,
655
],
"score": 1.0,
"content": "MVGRL (Hassani & Khasahmadi, 2020), DGI (Velickovic et al., 2019), BGRL (Thakoor et al.,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
652,
506,
667
],
"spans": [
{
"bbox": [
105,
652,
506,
667
],
"score": 1.0,
"content": "2021), CCA-SSG (Zhang et al., 2021), etc. All of them use graph convolutional networks (GCN) to",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
665,
211,
677
],
"spans": [
{
"bbox": [
105,
665,
211,
677
],
"score": 1.0,
"content": "create graph embeddings.",
"type": "text"
}
],
"index": 47
}
],
"index": 41.5
},
{
"type": "text",
"bbox": [
108,
687,
504,
732
],
"lines": [
{
"bbox": [
106,
687,
505,
700
],
"spans": [
{
"bbox": [
106,
687,
505,
700
],
"score": 1.0,
"content": "Augmentation-free node-level GCL A general problem with domain-agnostic graph augmenta-",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
699,
506,
712
],
"spans": [
{
"bbox": [
105,
699,
506,
712
],
"score": 1.0,
"content": "tions is that they can have unpredictable effects on graph semantics (Trivedi et al., 2022), as even",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
710,
505,
722
],
"spans": [
{
"bbox": [
105,
710,
505,
722
],
"score": 1.0,
"content": "minor augmentations can potentially result in a semantically different graph. This motivated devel-",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
721,
506,
733
],
"spans": [
{
"bbox": [
105,
721,
506,
733
],
"score": 1.0,
"content": "opment of augmentation-free GCL methods. Here positive pairs are pairs of nodes that are located",
"type": "text"
}
],
"index": 51
}
],
"index": 49.5
}
],
"page_idx": 1,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
763
],
"spans": [
{
"bbox": [
301,
750,
310,
763
],
"score": 1.0,
"content": "2",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
137
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
252,
95
],
"score": 1.0,
"content": "methods typically use node features",
"type": "text"
},
{
"bbox": [
252,
83,
262,
93
],
"score": 0.3,
"content": "\\mathbf { X }",
"type": "inline_equation"
},
{
"bbox": [
263,
83,
292,
95
],
"score": 1.0,
"content": "of size",
"type": "text"
},
{
"bbox": [
292,
83,
321,
93
],
"score": 0.91,
"content": "n \\times D",
"type": "inline_equation"
},
{
"bbox": [
321,
83,
348,
95
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
349,
83,
382,
94
],
"score": 0.93,
"content": "n = | \\mathcal { V } |",
"type": "inline_equation"
},
{
"bbox": [
383,
83,
505,
95
],
"score": 1.0,
"content": "and employ a neural network,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
104,
90,
507,
109
],
"spans": [
{
"bbox": [
104,
90,
418,
109
],
"score": 1.0,
"content": "usually a graph convolutional network (GCN) (Kipf & Welling, 2017), for the",
"type": "text"
},
{
"bbox": [
418,
93,
464,
104
],
"score": 0.91,
"content": "\\mathbf { R } ^ { D } \\to \\mathbf { R } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
465,
90,
507,
109
],
"score": 1.0,
"content": "mapping.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
102,
506,
118
],
"spans": [
{
"bbox": [
105,
102,
506,
118
],
"score": 1.0,
"content": "GCL methods also pull connected nodes together, sometimes explicitly through their loss function,",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
114,
505,
129
],
"spans": [
{
"bbox": [
105,
114,
505,
129
],
"score": 1.0,
"content": "but also implicitly through the GCN architecture (Trivedi et al., 2022; Wang et al., 2023; Guo et al.,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
124,
136,
140
],
"spans": [
{
"bbox": [
105,
124,
136,
140
],
"score": 1.0,
"content": "2023).",
"type": "text"
}
],
"index": 4
}
],
"index": 2,
"bbox_fs": [
104,
83,
507,
140
]
},
{
"type": "text",
"bbox": [
106,
143,
505,
221
],
"lines": [
{
"bbox": [
105,
142,
506,
156
],
"spans": [
{
"bbox": [
105,
142,
506,
156
],
"score": 1.0,
"content": "Recent work (Kruiger et al., 2017; Zhu et al., 2020a; Zhong et al., 2023; Bohm et al., 2022) pointed ¨",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
154,
505,
167
],
"spans": [
{
"bbox": [
105,
154,
476,
167
],
"score": 1.0,
"content": "out deep connections between graph layout and neighbor embedding algorithms such as",
"type": "text"
},
{
"bbox": [
477,
155,
482,
164
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
154,
505,
167
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
106,
165,
504,
178
],
"spans": [
{
"bbox": [
106,
165,
504,
178
],
"score": 1.0,
"content": "(Van der Maaten & Hinton, 2008) or UMAP (McInnes et al., 2018), which are based on neigh-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
177,
504,
188
],
"spans": [
{
"bbox": [
106,
177,
504,
188
],
"score": 1.0,
"content": "borhood preservation. In parallel, another line of work explored connections between neighbor",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
187,
505,
200
],
"spans": [
{
"bbox": [
105,
187,
505,
200
],
"score": 1.0,
"content": "embeddings and contrastive learning (Damrich et al., 2022; Bohm et al., 2023; Hu et al., 2023). ¨",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
198,
505,
211
],
"spans": [
{
"bbox": [
105,
198,
505,
211
],
"score": 1.0,
"content": "This raises the question to what extent neighbor embedding and contrastive neighbor embedding",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
208,
403,
223
],
"spans": [
{
"bbox": [
106,
208,
403,
223
],
"score": 1.0,
"content": "algorithms (see Section 3) can be useful for graph representation learning.",
"type": "text"
}
],
"index": 11
}
],
"index": 8,
"bbox_fs": [
105,
142,
506,
223
]
},
{
"type": "text",
"bbox": [
107,
226,
505,
292
],
"lines": [
{
"bbox": [
105,
226,
505,
239
],
"spans": [
{
"bbox": [
105,
226,
477,
239
],
"score": 1.0,
"content": "In this work, we answer this question. We introduce a novel graph layout algorithm, graph",
"type": "text"
},
{
"bbox": [
477,
227,
482,
236
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
226,
505,
239
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
237,
505,
249
],
"spans": [
{
"bbox": [
105,
237,
505,
249
],
"score": 1.0,
"content": "(Figure 1), and show that it strongly outperforms existing methods. We also introduce a novel,",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
247,
506,
261
],
"spans": [
{
"bbox": [
105,
247,
506,
261
],
"score": 1.0,
"content": "augmentation-free, GCL algorithm, graph CNE (Figure 1), based on the framework for contrastive",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
106,
259,
505,
271
],
"spans": [
{
"bbox": [
106,
259,
505,
271
],
"score": 1.0,
"content": "neighbor embeddings, and show that it reaches competitive GCL performance without using GCNs.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
268,
506,
284
],
"spans": [
{
"bbox": [
105,
268,
506,
284
],
"score": 1.0,
"content": "Conceptually, we present a single coherent framework for node-level graph representation learning,",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
106,
281,
440,
294
],
"spans": [
{
"bbox": [
106,
281,
440,
294
],
"score": 1.0,
"content": "tying together graph layouts, graph contrastive learning, and neighbor embeddings.",
"type": "text"
}
],
"index": 17
}
],
"index": 14.5,
"bbox_fs": [
105,
226,
506,
294
]
},
{
"type": "title",
"bbox": [
108,
307,
211,
320
],
"lines": [
{
"bbox": [
104,
306,
213,
323
],
"spans": [
{
"bbox": [
104,
306,
213,
323
],
"score": 1.0,
"content": "2 RELATED WORK",
"type": "text"
}
],
"index": 18
}
],
"index": 18
},
{
"type": "text",
"bbox": [
107,
332,
505,
421
],
"lines": [
{
"bbox": [
106,
333,
505,
345
],
"spans": [
{
"bbox": [
106,
333,
505,
345
],
"score": 1.0,
"content": "Graph layouts Graph layout algorithms have traditionally been based on spring models, where",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
343,
505,
356
],
"spans": [
{
"bbox": [
105,
343,
332,
356
],
"score": 1.0,
"content": "every connected pair of nodes feels an attractive force",
"type": "text"
},
{
"bbox": [
333,
344,
345,
355
],
"score": 0.9,
"content": "F _ { a }",
"type": "inline_equation"
},
{
"bbox": [
345,
343,
505,
356
],
"score": 1.0,
"content": "and all pairs of nodes feel a repulsive",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
104,
353,
504,
370
],
"spans": [
{
"bbox": [
104,
353,
129,
370
],
"score": 1.0,
"content": "force",
"type": "text"
},
{
"bbox": [
129,
355,
141,
366
],
"score": 0.86,
"content": "F _ { r }",
"type": "inline_equation"
},
{
"bbox": [
142,
353,
409,
370
],
"score": 1.0,
"content": "(force-directed graph layouts). Many algorithms can be written as",
"type": "text"
},
{
"bbox": [
410,
355,
447,
368
],
"score": 0.92,
"content": "F _ { a } = d _ { i j } ^ { a }",
"type": "inline_equation"
},
{
"bbox": [
448,
353,
466,
370
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
466,
355,
504,
368
],
"score": 0.92,
"content": "F _ { r } = d _ { i j } ^ { r }",
"type": "inline_equation"
}
],
"index": 21
},
{
"bbox": [
105,
365,
505,
379
],
"spans": [
{
"bbox": [
105,
365,
198,
379
],
"score": 1.0,
"content": "(Noack, 2007), where",
"type": "text"
},
{
"bbox": [
198,
366,
212,
378
],
"score": 0.9,
"content": "d _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
212,
365,
505,
379
],
"score": 1.0,
"content": "is the embedding distance between nodes. For example, Fruchterman–",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
377,
505,
389
],
"spans": [
{
"bbox": [
105,
377,
296,
389
],
"score": 1.0,
"content": "Reingold algorithm, also known as FDP, uses",
"type": "text"
},
{
"bbox": [
297,
377,
363,
388
],
"score": 0.92,
"content": "a = 2 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
363,
377,
505,
389
],
"score": 1.0,
"content": "(Fruchterman & Reingold, 1991);",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
387,
506,
401
],
"spans": [
{
"bbox": [
105,
387,
177,
401
],
"score": 1.0,
"content": "ForceAtlas2 uses",
"type": "text"
},
{
"bbox": [
178,
388,
238,
399
],
"score": 0.92,
"content": "a = 1 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
238,
387,
380,
401
],
"score": 1.0,
"content": "(Jacomy et al., 2014); LinLog uses",
"type": "text"
},
{
"bbox": [
381,
388,
441,
399
],
"score": 0.93,
"content": "a = 0 , r = - 1",
"type": "inline_equation"
},
{
"bbox": [
441,
387,
506,
401
],
"score": 1.0,
"content": "(Noack, 2007).",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
398,
506,
411
],
"spans": [
{
"bbox": [
105,
398,
506,
411
],
"score": 1.0,
"content": "Efficient implementations can be based on Barnes–Hut approximation of the repulsive forces, as in",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
106,
410,
490,
422
],
"spans": [
{
"bbox": [
106,
410,
490,
422
],
"score": 1.0,
"content": "SFDP (Hu, 2005). Relationship to neighbour embeddings was discussed by Bohm et al. (2022). ¨",
"type": "text"
}
],
"index": 26
}
],
"index": 22.5,
"bbox_fs": [
104,
333,
506,
422
]
},
{
"type": "text",
"bbox": [
107,
432,
505,
532
],
"lines": [
{
"bbox": [
106,
433,
505,
446
],
"spans": [
{
"bbox": [
106,
433,
225,
446
],
"score": 1.0,
"content": "Graph layouts inspired by",
"type": "text"
},
{
"bbox": [
226,
434,
231,
443
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
231,
433,
505,
446
],
"score": 1.0,
"content": "-SNE Several recent graph layout algorithms have been inspired",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
443,
506,
457
],
"spans": [
{
"bbox": [
105,
443,
450,
457
],
"score": 1.0,
"content": "by neighbor embeddings. tsNET (Kruiger et al., 2017) applied modified version of",
"type": "text"
},
{
"bbox": [
450,
444,
455,
454
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
455,
443,
506,
457
],
"score": 1.0,
"content": "-SNE to the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
104,
454,
506,
468
],
"spans": [
{
"bbox": [
104,
454,
506,
468
],
"score": 1.0,
"content": "pairwise shortest path distances between all nodes. DRGraph (Zhu et al., 2020a) made tsNET faster",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
465,
504,
479
],
"spans": [
{
"bbox": [
105,
465,
309,
479
],
"score": 1.0,
"content": "by using negative sampling (Mikolov et al., 2013).",
"type": "text"
},
{
"bbox": [
309,
466,
315,
476
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
315,
465,
491,
479
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023) suggested custom",
"type": "text"
},
{
"bbox": [
491,
466,
504,
477
],
"score": 0.88,
"content": "F _ { a }",
"type": "inline_equation"
}
],
"index": 30
},
{
"bbox": [
105,
477,
505,
490
],
"spans": [
{
"bbox": [
105,
477,
123,
490
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
123,
477,
135,
488
],
"score": 0.88,
"content": "F _ { r }",
"type": "inline_equation"
},
{
"bbox": [
135,
477,
209,
490
],
"score": 1.0,
"content": "forces inspired by",
"type": "text"
},
{
"bbox": [
209,
478,
214,
487
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
214,
477,
505,
490
],
"score": 1.0,
"content": "-SNE and adopted interpolation-based approximation of Linderman et al.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
486,
505,
502
],
"spans": [
{
"bbox": [
105,
486,
284,
502
],
"score": 1.0,
"content": "(2019). Below we will show that our graph",
"type": "text"
},
{
"bbox": [
284,
488,
289,
497
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
289,
486,
443,
502
],
"score": 1.0,
"content": "-SNE outperforms both DRGraph and",
"type": "text"
},
{
"bbox": [
443,
488,
448,
498
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
448,
486,
505,
502
],
"score": 1.0,
"content": "-FDP. Finally,",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
499,
505,
511
],
"spans": [
{
"bbox": [
106,
499,
346,
511
],
"score": 1.0,
"content": "Leow et al. (2019) also suggested an algorithm called ‘graph",
"type": "text"
},
{
"bbox": [
346,
500,
351,
509
],
"score": 0.57,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
351,
499,
505,
511
],
"score": 1.0,
"content": "-SNE’, that used a graph convolutional",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
508,
506,
524
],
"spans": [
{
"bbox": [
104,
508,
477,
524
],
"score": 1.0,
"content": "network (Kipf & Welling, 2017) to build a parametric mapping optimizing a combination of",
"type": "text"
},
{
"bbox": [
477,
510,
482,
520
],
"score": 0.73,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
482,
508,
506,
524
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
520,
505,
533
],
"spans": [
{
"bbox": [
105,
520,
474,
533
],
"score": 1.0,
"content": "losses on node features and on shortest graph distances; it has almost no relation to our graph",
"type": "text"
},
{
"bbox": [
474,
522,
479,
530
],
"score": 0.65,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
480,
520,
505,
533
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 35
}
],
"index": 31,
"bbox_fs": [
104,
433,
506,
533
]
},
{
"type": "text",
"bbox": [
107,
543,
505,
676
],
"lines": [
{
"bbox": [
105,
543,
505,
557
],
"spans": [
{
"bbox": [
105,
543,
505,
557
],
"score": 1.0,
"content": "Node-level graph contrastive learning The basic principle behind contrastive learning is to learn",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
555,
505,
568
],
"spans": [
{
"bbox": [
105,
555,
505,
568
],
"score": 1.0,
"content": "data representation by contrasting pairs of observations that are similar to each other (positive pairs)",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
565,
505,
579
],
"spans": [
{
"bbox": [
105,
565,
505,
579
],
"score": 1.0,
"content": "with those that are dissimilar to each other (negative pairs). In computer vision, positive pairs are",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
104,
576,
506,
591
],
"spans": [
{
"bbox": [
104,
576,
506,
591
],
"score": 1.0,
"content": "generated via data augmentation, e.g. in SimCLR (Chen et al., 2020). Graph contrastive learning",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
588,
505,
600
],
"spans": [
{
"bbox": [
105,
588,
505,
600
],
"score": 1.0,
"content": "(GCL) requires node features (as input to the network) and can be graph-level or node-level, de-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
599,
505,
612
],
"spans": [
{
"bbox": [
105,
599,
505,
612
],
"score": 1.0,
"content": "pending on whether representations are obtained for a set of graphs or for the set of nodes of a single",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
104,
609,
506,
623
],
"spans": [
{
"bbox": [
104,
609,
506,
623
],
"score": 1.0,
"content": "graph. Graph-level GCL is based on graph augmentations, such as node dropping or edge pertur-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
620,
506,
633
],
"spans": [
{
"bbox": [
105,
620,
506,
633
],
"score": 1.0,
"content": "bation, e.g. in GraphCL (You et al., 2020). Prominent examples of node-level GCL algorithms that",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
632,
505,
644
],
"spans": [
{
"bbox": [
105,
632,
505,
644
],
"score": 1.0,
"content": "are also based on graph augmentations include GRACE (Zhu et al., 2020b), GCA (Zhu et al., 2021),",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
641,
505,
655
],
"spans": [
{
"bbox": [
105,
641,
505,
655
],
"score": 1.0,
"content": "MVGRL (Hassani & Khasahmadi, 2020), DGI (Velickovic et al., 2019), BGRL (Thakoor et al.,",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
652,
506,
667
],
"spans": [
{
"bbox": [
105,
652,
506,
667
],
"score": 1.0,
"content": "2021), CCA-SSG (Zhang et al., 2021), etc. All of them use graph convolutional networks (GCN) to",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
105,
665,
211,
677
],
"spans": [
{
"bbox": [
105,
665,
211,
677
],
"score": 1.0,
"content": "create graph embeddings.",
"type": "text"
}
],
"index": 47
}
],
"index": 41.5,
"bbox_fs": [
104,
543,
506,
677
]
},
{
"type": "text",
"bbox": [
108,
687,
504,
732
],
"lines": [
{
"bbox": [
106,
687,
505,
700
],
"spans": [
{
"bbox": [
106,
687,
505,
700
],
"score": 1.0,
"content": "Augmentation-free node-level GCL A general problem with domain-agnostic graph augmenta-",
"type": "text"
}
],
"index": 48
},
{
"bbox": [
105,
699,
506,
712
],
"spans": [
{
"bbox": [
105,
699,
506,
712
],
"score": 1.0,
"content": "tions is that they can have unpredictable effects on graph semantics (Trivedi et al., 2022), as even",
"type": "text"
}
],
"index": 49
},
{
"bbox": [
105,
710,
505,
722
],
"spans": [
{
"bbox": [
105,
710,
505,
722
],
"score": 1.0,
"content": "minor augmentations can potentially result in a semantically different graph. This motivated devel-",
"type": "text"
}
],
"index": 50
},
{
"bbox": [
105,
721,
506,
733
],
"spans": [
{
"bbox": [
105,
721,
506,
733
],
"score": 1.0,
"content": "opment of augmentation-free GCL methods. Here positive pairs are pairs of nodes that are located",
"type": "text"
}
],
"index": 51
},
{
"bbox": [
106,
82,
505,
94
],
"spans": [
{
"bbox": [
106,
82,
505,
94
],
"score": 1.0,
"content": "close to each other in terms of graph distance. AFGRL (Lee et al., 2022) and AF-GCL (Li et al.,",
"type": "text",
"cross_page": true
}
],
"index": 0
},
{
"bbox": [
106,
93,
504,
105
],
"spans": [
{
"bbox": [
106,
93,
437,
105
],
"score": 1.0,
"content": "2023) treat nodes with small shortest path distance as candidate positives, and use",
"type": "text",
"cross_page": true
},
{
"bbox": [
438,
94,
445,
104
],
"score": 0.76,
"content": "k",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
445,
93,
504,
105
],
"score": 1.0,
"content": "nearest neigh-",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
105,
104,
505,
117
],
"spans": [
{
"bbox": [
105,
104,
505,
117
],
"score": 1.0,
"content": "bors in GCN-based node representations to select actual positives. Local-GCL (Zhang et al., 2022)",
"type": "text",
"cross_page": true
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
128
],
"spans": [
{
"bbox": [
105,
115,
505,
128
],
"score": 1.0,
"content": "uses all first-order graph neighbors as positives, and employs random Fourier features to approxi-",
"type": "text",
"cross_page": true
}
],
"index": 3
},
{
"bbox": [
105,
126,
471,
139
],
"spans": [
{
"bbox": [
105,
126,
128,
139
],
"score": 1.0,
"content": "mate",
"type": "text",
"cross_page": true
},
{
"bbox": [
129,
126,
156,
138
],
"score": 0.92,
"content": "\\mathcal { O } ( n ^ { 2 } )",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
156,
126,
471,
139
],
"score": 1.0,
"content": "repulsive forces. All of these methods are also based on the GCN architecture.",
"type": "text",
"cross_page": true
}
],
"index": 4
}
],
"index": 49.5,
"bbox_fs": [
105,
687,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
138
],
"lines": [
{
"bbox": [
106,
82,
505,
94
],
"spans": [
{
"bbox": [
106,
82,
505,
94
],
"score": 1.0,
"content": "close to each other in terms of graph distance. AFGRL (Lee et al., 2022) and AF-GCL (Li et al.,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
93,
504,
105
],
"spans": [
{
"bbox": [
106,
93,
437,
105
],
"score": 1.0,
"content": "2023) treat nodes with small shortest path distance as candidate positives, and use",
"type": "text"
},
{
"bbox": [
438,
94,
445,
104
],
"score": 0.76,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
445,
93,
504,
105
],
"score": 1.0,
"content": "nearest neigh-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
104,
505,
117
],
"spans": [
{
"bbox": [
105,
104,
505,
117
],
"score": 1.0,
"content": "bors in GCN-based node representations to select actual positives. Local-GCL (Zhang et al., 2022)",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
115,
505,
128
],
"spans": [
{
"bbox": [
105,
115,
505,
128
],
"score": 1.0,
"content": "uses all first-order graph neighbors as positives, and employs random Fourier features to approxi-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
126,
471,
139
],
"spans": [
{
"bbox": [
105,
126,
128,
139
],
"score": 1.0,
"content": "mate",
"type": "text"
},
{
"bbox": [
129,
126,
156,
138
],
"score": 0.92,
"content": "\\mathcal { O } ( n ^ { 2 } )",
"type": "inline_equation"
},
{
"bbox": [
156,
126,
471,
139
],
"score": 1.0,
"content": "repulsive forces. All of these methods are also based on the GCN architecture.",
"type": "text"
}
],
"index": 4
}
],
"index": 2
},
{
"type": "title",
"bbox": [
108,
154,
200,
167
],
"lines": [
{
"bbox": [
104,
152,
202,
170
],
"spans": [
{
"bbox": [
104,
152,
202,
170
],
"score": 1.0,
"content": "3 BACKGROUND",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "title",
"bbox": [
108,
179,
238,
190
],
"lines": [
{
"bbox": [
106,
179,
239,
191
],
"spans": [
{
"bbox": [
106,
179,
239,
191
],
"score": 1.0,
"content": "3.1 NEIGHBOR EMBEDDINGS",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "text",
"bbox": [
106,
199,
505,
255
],
"lines": [
{
"bbox": [
106,
200,
504,
212
],
"spans": [
{
"bbox": [
106,
200,
371,
212
],
"score": 1.0,
"content": "Neighbor embeddings are a family of methods aiming to embed",
"type": "text"
},
{
"bbox": [
372,
202,
379,
210
],
"score": 0.71,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
380,
200,
504,
212
],
"score": 1.0,
"content": "observations from some high-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
104,
209,
505,
224
],
"spans": [
{
"bbox": [
104,
209,
212,
224
],
"score": 1.0,
"content": "dimensional metric space",
"type": "text"
},
{
"bbox": [
212,
212,
222,
221
],
"score": 0.79,
"content": "\\mathcal { X }",
"type": "inline_equation"
},
{
"bbox": [
222,
209,
488,
224
],
"score": 1.0,
"content": "into a lower-dimensional (usually two-dimensional) vector space",
"type": "text"
},
{
"bbox": [
488,
210,
501,
221
],
"score": 0.87,
"content": "\\bar { \\mathbb { R } } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
501,
209,
505,
224
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
104,
220,
506,
236
],
"spans": [
{
"bbox": [
104,
220,
506,
236
],
"score": 1.0,
"content": "such that neighborhood relationships between observations are preserved in the embedding space.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
232,
503,
245
],
"spans": [
{
"bbox": [
106,
232,
147,
245
],
"score": 1.0,
"content": "Typically,",
"type": "text"
},
{
"bbox": [
148,
233,
158,
243
],
"score": 0.81,
"content": "\\mathcal { X }",
"type": "inline_equation"
},
{
"bbox": [
158,
232,
270,
245
],
"score": 1.0,
"content": "is another real-valued space",
"type": "text"
},
{
"bbox": [
270,
233,
282,
243
],
"score": 0.87,
"content": "\\mathbb { R } ^ { p }",
"type": "inline_equation"
},
{
"bbox": [
283,
232,
306,
245
],
"score": 1.0,
"content": ", with",
"type": "text"
},
{
"bbox": [
306,
233,
333,
244
],
"score": 0.91,
"content": "d \\ll p",
"type": "inline_equation"
},
{
"bbox": [
333,
232,
469,
245
],
"score": 1.0,
"content": ". We denote the original vectors as",
"type": "text"
},
{
"bbox": [
469,
233,
503,
244
],
"score": 0.92,
"content": "\\mathbf { x } _ { i } \\in \\mathbb { R } ^ { p }",
"type": "inline_equation"
}
],
"index": 10
},
{
"bbox": [
105,
243,
265,
256
],
"spans": [
{
"bbox": [
105,
243,
226,
256
],
"score": 1.0,
"content": "and the embedding vectors as",
"type": "text"
},
{
"bbox": [
227,
243,
261,
255
],
"score": 0.92,
"content": "\\mathbf { y } _ { i } \\in \\bar { \\mathbb { R } } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
261,
243,
265,
256
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 11
}
],
"index": 9
},
{
"type": "text",
"bbox": [
107,
260,
504,
305
],
"lines": [
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
327,
273
],
"score": 1.0,
"content": "One of the most popular neighbor embedding methods,",
"type": "text"
},
{
"bbox": [
327,
262,
333,
271
],
"score": 0.76,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
333,
259,
506,
273
],
"score": 1.0,
"content": "-distributed stochastic neighbor embedding",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
270,
506,
285
],
"spans": [
{
"bbox": [
105,
270,
110,
285
],
"score": 1.0,
"content": "(",
"type": "text"
},
{
"bbox": [
110,
272,
114,
282
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
115,
270,
210,
285
],
"score": 1.0,
"content": "-SNE; Van der Maaten",
"type": "text"
},
{
"bbox": [
211,
272,
221,
282
],
"score": 0.5,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
221,
270,
506,
285
],
"score": 1.0,
"content": "Hinton, 2008) is an extension of the stochastic neighbor embedding",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
283,
506,
295
],
"spans": [
{
"bbox": [
106,
283,
344,
295
],
"score": 1.0,
"content": "(SNE) originally suggested by Hinton & Roweis (2002).",
"type": "text"
},
{
"bbox": [
344,
283,
349,
293
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
349,
283,
506,
295
],
"score": 1.0,
"content": "-SNE minimizes the Kullback-Leibler",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
292,
446,
308
],
"spans": [
{
"bbox": [
105,
292,
397,
308
],
"score": 1.0,
"content": "divergence between the high-dimensional and low-dimensional affinities",
"type": "text"
},
{
"bbox": [
397,
295,
411,
306
],
"score": 0.87,
"content": "p _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
411,
292,
429,
308
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
429,
295,
441,
306
],
"score": 0.88,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
442,
292,
446,
308
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5
},
{
"type": "interline_equation",
"bbox": [
233,
311,
378,
339
],
"lines": [
{
"bbox": [
233,
311,
378,
339
],
"spans": [
{
"bbox": [
233,
311,
378,
339
],
"score": 0.95,
"content": "\\mathcal { L } = \\mathrm { { K L } } ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i j } p _ { i j } \\log \\frac { p _ { i j } } { q _ { i j } } .",
"type": "interline_equation",
"image_path": "84f384e8029d2b4f62221afe7a55436728d67f8a5b0edba25f989a97d956da3b.jpg"
}
]
}
],
"index": 16.5,
"virtual_lines": [
{
"bbox": [
233,
311,
378,
325.0
],
"spans": [],
"index": 16
},
{
"bbox": [
233,
325.0,
378,
339.0
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
106,
344,
505,
378
],
"lines": [
{
"bbox": [
105,
344,
505,
357
],
"spans": [
{
"bbox": [
105,
344,
505,
357
],
"score": 1.0,
"content": "Both affinity matrices are defined to be symmetric, positive, and to sum to 1. In the original algo-",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
356,
505,
368
],
"spans": [
{
"bbox": [
106,
356,
132,
368
],
"score": 1.0,
"content": "rithm,",
"type": "text"
},
{
"bbox": [
133,
356,
142,
366
],
"score": 0.62,
"content": "\\mathbf { P }",
"type": "inline_equation"
},
{
"bbox": [
142,
356,
505,
368
],
"score": 1.0,
"content": "was computed using adaptive Gaussian kernels, but almost the same results can be obtained",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
367,
500,
379
],
"spans": [
{
"bbox": [
105,
367,
288,
379
],
"score": 1.0,
"content": "simply by normalizing and symmetrizing the",
"type": "text"
},
{
"bbox": [
288,
367,
310,
377
],
"score": 0.74,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
310,
367,
500,
379
],
"score": 1.0,
"content": "graph adjacency matrix A (Bohm et al., 2022): ¨",
"type": "text"
}
],
"index": 20
}
],
"index": 19
},
{
"type": "interline_equation",
"bbox": [
262,
383,
349,
410
],
"lines": [
{
"bbox": [
262,
383,
349,
410
],
"spans": [
{
"bbox": [
262,
383,
349,
410
],
"score": 0.96,
"content": "\\mathbf { P } = { \\frac { \\mathbf { A } / k + \\mathbf { A } ^ { \\top } / k } { 2 n } } .",
"type": "interline_equation",
"image_path": "f689fa8500db6ec61882ece7b6b97aba88c5d29597b200782c6426f1484c3416.jpg"
}
]
}
],
"index": 21.5,
"virtual_lines": [
{
"bbox": [
262,
383,
349,
396.5
],
"spans": [],
"index": 21
},
{
"bbox": [
262,
396.5,
349,
410.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
107,
414,
504,
447
],
"lines": [
{
"bbox": [
105,
414,
504,
427
],
"spans": [
{
"bbox": [
105,
414,
193,
427
],
"score": 1.0,
"content": "Here A has element",
"type": "text"
},
{
"bbox": [
194,
415,
229,
426
],
"score": 0.91,
"content": "a _ { i j } = 1",
"type": "inline_equation"
},
{
"bbox": [
229,
414,
240,
427
],
"score": 1.0,
"content": "if",
"type": "text"
},
{
"bbox": [
241,
415,
252,
426
],
"score": 0.86,
"content": "\\mathbf { x } _ { j }",
"type": "inline_equation"
},
{
"bbox": [
253,
414,
293,
427
],
"score": 1.0,
"content": "is within",
"type": "text"
},
{
"bbox": [
294,
415,
300,
424
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
301,
414,
389,
427
],
"score": 1.0,
"content": "nearest neighbors of",
"type": "text"
},
{
"bbox": [
389,
416,
399,
425
],
"score": 0.85,
"content": "\\mathbf { x } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
399,
414,
497,
427
],
"score": 1.0,
"content": ". Reasonable values of",
"type": "text"
},
{
"bbox": [
497,
415,
504,
424
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
}
],
"index": 23
},
{
"bbox": [
105,
425,
506,
438
],
"spans": [
{
"bbox": [
105,
425,
361,
438
],
"score": 1.0,
"content": "typically lie between 10 and 100. Low-dimensional affinities",
"type": "text"
},
{
"bbox": [
361,
425,
372,
437
],
"score": 0.8,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
372,
425,
433,
438
],
"score": 1.0,
"content": "are defined in",
"type": "text"
},
{
"bbox": [
433,
426,
438,
435
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
438,
425,
495,
438
],
"score": 1.0,
"content": "-SNE using a",
"type": "text"
},
{
"bbox": [
496,
426,
501,
435
],
"score": 0.73,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
501,
425,
506,
438
],
"score": 1.0,
"content": "-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
436,
433,
448
],
"spans": [
{
"bbox": [
106,
436,
433,
448
],
"score": 1.0,
"content": "distribution kernel with one degree of freedom, also known as the Cauchy kernel:",
"type": "text"
}
],
"index": 25
}
],
"index": 24
},
{
"type": "interline_equation",
"bbox": [
236,
452,
375,
483
],
"lines": [
{
"bbox": [
236,
452,
375,
483
],
"spans": [
{
"bbox": [
236,
452,
375,
483
],
"score": 0.95,
"content": "q _ { i j } = \\frac { ( 1 + \\| \\mathbf { y } _ { i } - \\mathbf { y } _ { j } \\| ^ { 2 } ) ^ { - 1 } } { \\sum _ { k \\neq l } ( 1 + \\| \\mathbf { y } _ { l } - \\mathbf { y } _ { k } \\| ^ { 2 } ) ^ { - 1 } } .",
"type": "interline_equation",
"image_path": "c46e8a49680d153032d10a818f4f19b1812d9912d8c69d0757d68bfeb6234671.jpg"
}
]
}
],
"index": 26.5,
"virtual_lines": [
{
"bbox": [
236,
452,
375,
467.5
],
"spans": [],
"index": 26
},
{
"bbox": [
236,
467.5,
375,
483.0
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
487,
504,
510
],
"lines": [
{
"bbox": [
106,
487,
504,
500
],
"spans": [
{
"bbox": [
106,
487,
504,
500
],
"score": 1.0,
"content": "The original SNE algorithm used Gaussian kernel instead of Cauchy, which led to worse results",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
498,
353,
510
],
"spans": [
{
"bbox": [
106,
498,
353,
510
],
"score": 1.0,
"content": "when embedding high-dimensional data (Kobak et al., 2019).",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5
},
{
"type": "text",
"bbox": [
106,
514,
505,
582
],
"lines": [
{
"bbox": [
106,
515,
505,
528
],
"spans": [
{
"bbox": [
106,
515,
297,
528
],
"score": 1.0,
"content": "Even though it is usually not presented like that,",
"type": "text"
},
{
"bbox": [
297,
516,
302,
525
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
303,
515,
505,
528
],
"score": 1.0,
"content": "-SNE can be thought of as a graph layout algorithm",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
526,
505,
539
],
"spans": [
{
"bbox": [
106,
526,
120,
539
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
121,
527,
142,
537
],
"score": 0.55,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
143,
526,
505,
539
],
"score": 1.0,
"content": "graphs, in particular after the reformulation in Equation 2. During optimization, neighbor-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
537,
505,
550
],
"spans": [
{
"bbox": [
105,
537,
505,
550
],
"score": 1.0,
"content": "ing nodes (sharing an edge) feel attraction, whereas all nodes feel repulsion, arising through the",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
548,
505,
561
],
"spans": [
{
"bbox": [
105,
548,
273,
561
],
"score": 1.0,
"content": "normalization in Equation 3. In practice,",
"type": "text"
},
{
"bbox": [
273,
549,
279,
558
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
279,
548,
505,
561
],
"score": 1.0,
"content": "-SNE optimization can be accelerated by an approxima-",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
558,
506,
573
],
"spans": [
{
"bbox": [
105,
558,
506,
573
],
"score": 1.0,
"content": "tion of the repulsive force field based on the Barnes–Hut algorithm (Van Der Maaten, 2014; Yang",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
570,
335,
582
],
"spans": [
{
"bbox": [
106,
570,
335,
582
],
"score": 1.0,
"content": "et al., 2013) or on interpolation (Linderman et al., 2019).",
"type": "text"
}
],
"index": 35
}
],
"index": 32.5
},
{
"type": "title",
"bbox": [
109,
595,
302,
606
],
"lines": [
{
"bbox": [
106,
595,
302,
608
],
"spans": [
{
"bbox": [
106,
595,
302,
608
],
"score": 1.0,
"content": "3.2 CONTRASTIVE NEIGHBOR EMBEDDINGS",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
107,
615,
505,
704
],
"lines": [
{
"bbox": [
105,
615,
506,
629
],
"spans": [
{
"bbox": [
105,
615,
506,
629
],
"score": 1.0,
"content": "The contrastive neighbor embedding (CNE) algorithm (Damrich et al., 2022) is a flexible framework",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
627,
505,
639
],
"spans": [
{
"bbox": [
105,
627,
210,
639
],
"score": 1.0,
"content": "that also operates on the",
"type": "text"
},
{
"bbox": [
210,
627,
232,
637
],
"score": 0.34,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
233,
627,
505,
639
],
"score": 1.0,
"content": "graph of the data, and optimizes the embedding in order to place",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
637,
506,
650
],
"spans": [
{
"bbox": [
105,
637,
506,
650
],
"score": 1.0,
"content": "connected nodes closer together than unconnected pairs of nodes. Damrich et al. (2022) considered",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
648,
505,
661
],
"spans": [
{
"bbox": [
105,
648,
505,
661
],
"score": 1.0,
"content": "three different loss functions: NCE (noise-contrastive estimation) (Gutmann & Hyvarinen, 2010), ¨",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
659,
505,
673
],
"spans": [
{
"bbox": [
105,
659,
505,
673
],
"score": 1.0,
"content": "InfoNCE (Jozefowicz et al., 2016; Oord et al., 2018), and negative sampling (Mikolov et al., 2013).",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
670,
505,
684
],
"spans": [
{
"bbox": [
105,
670,
505,
684
],
"score": 1.0,
"content": "These loss functions are called contrastive because they are based on contrasting edges and non-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
680,
506,
695
],
"spans": [
{
"bbox": [
105,
680,
506,
695
],
"score": 1.0,
"content": "edges in the same mini-batch, and do not require a global normalization like in Equation 3. Using",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
692,
304,
705
],
"spans": [
{
"bbox": [
105,
692,
273,
705
],
"score": 1.0,
"content": "NCE and InfoNCE in CNE approximates",
"type": "text"
},
{
"bbox": [
274,
694,
278,
703
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
279,
692,
304,
705
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 44
}
],
"index": 40.5
},
{
"type": "text",
"bbox": [
107,
709,
504,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Damrich et al. (2022) also considered parametric embeddings, where a neural network (usually a",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
381,
734
],
"score": 1.0,
"content": "fully-connected network) is trained to produce embedding vectors",
"type": "text"
},
{
"bbox": [
381,
720,
433,
732
],
"score": 0.93,
"content": "\\mathbf { y } _ { i } ~ = ~ f ( \\mathbf { x } _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
433,
720,
505,
734
],
"score": 1.0,
"content": "using one of the",
"type": "text"
}
],
"index": 46
}
],
"index": 45.5
}
],
"page_idx": 2,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
25,
308,
38
],
"spans": [
{
"bbox": [
107,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
309,
760
],
"lines": [
{
"bbox": [
301,
750,
310,
762
],
"spans": [
{
"bbox": [
301,
750,
310,
762
],
"score": 1.0,
"content": "3",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "text",
"bbox": [
107,
82,
505,
138
],
"lines": [],
"index": 2,
"bbox_fs": [
105,
82,
505,
139
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
108,
154,
200,
167
],
"lines": [
{
"bbox": [
104,
152,
202,
170
],
"spans": [
{
"bbox": [
104,
152,
202,
170
],
"score": 1.0,
"content": "3 BACKGROUND",
"type": "text"
}
],
"index": 5
}
],
"index": 5
},
{
"type": "title",
"bbox": [
108,
179,
238,
190
],
"lines": [
{
"bbox": [
106,
179,
239,
191
],
"spans": [
{
"bbox": [
106,
179,
239,
191
],
"score": 1.0,
"content": "3.1 NEIGHBOR EMBEDDINGS",
"type": "text"
}
],
"index": 6
}
],
"index": 6
},
{
"type": "text",
"bbox": [
106,
199,
505,
255
],
"lines": [
{
"bbox": [
106,
200,
504,
212
],
"spans": [
{
"bbox": [
106,
200,
371,
212
],
"score": 1.0,
"content": "Neighbor embeddings are a family of methods aiming to embed",
"type": "text"
},
{
"bbox": [
372,
202,
379,
210
],
"score": 0.71,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
380,
200,
504,
212
],
"score": 1.0,
"content": "observations from some high-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
104,
209,
505,
224
],
"spans": [
{
"bbox": [
104,
209,
212,
224
],
"score": 1.0,
"content": "dimensional metric space",
"type": "text"
},
{
"bbox": [
212,
212,
222,
221
],
"score": 0.79,
"content": "\\mathcal { X }",
"type": "inline_equation"
},
{
"bbox": [
222,
209,
488,
224
],
"score": 1.0,
"content": "into a lower-dimensional (usually two-dimensional) vector space",
"type": "text"
},
{
"bbox": [
488,
210,
501,
221
],
"score": 0.87,
"content": "\\bar { \\mathbb { R } } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
501,
209,
505,
224
],
"score": 1.0,
"content": ",",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
104,
220,
506,
236
],
"spans": [
{
"bbox": [
104,
220,
506,
236
],
"score": 1.0,
"content": "such that neighborhood relationships between observations are preserved in the embedding space.",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
232,
503,
245
],
"spans": [
{
"bbox": [
106,
232,
147,
245
],
"score": 1.0,
"content": "Typically,",
"type": "text"
},
{
"bbox": [
148,
233,
158,
243
],
"score": 0.81,
"content": "\\mathcal { X }",
"type": "inline_equation"
},
{
"bbox": [
158,
232,
270,
245
],
"score": 1.0,
"content": "is another real-valued space",
"type": "text"
},
{
"bbox": [
270,
233,
282,
243
],
"score": 0.87,
"content": "\\mathbb { R } ^ { p }",
"type": "inline_equation"
},
{
"bbox": [
283,
232,
306,
245
],
"score": 1.0,
"content": ", with",
"type": "text"
},
{
"bbox": [
306,
233,
333,
244
],
"score": 0.91,
"content": "d \\ll p",
"type": "inline_equation"
},
{
"bbox": [
333,
232,
469,
245
],
"score": 1.0,
"content": ". We denote the original vectors as",
"type": "text"
},
{
"bbox": [
469,
233,
503,
244
],
"score": 0.92,
"content": "\\mathbf { x } _ { i } \\in \\mathbb { R } ^ { p }",
"type": "inline_equation"
}
],
"index": 10
},
{
"bbox": [
105,
243,
265,
256
],
"spans": [
{
"bbox": [
105,
243,
226,
256
],
"score": 1.0,
"content": "and the embedding vectors as",
"type": "text"
},
{
"bbox": [
227,
243,
261,
255
],
"score": 0.92,
"content": "\\mathbf { y } _ { i } \\in \\bar { \\mathbb { R } } ^ { d }",
"type": "inline_equation"
},
{
"bbox": [
261,
243,
265,
256
],
"score": 1.0,
"content": ".",
"type": "text"
}
],
"index": 11
}
],
"index": 9,
"bbox_fs": [
104,
200,
506,
256
]
},
{
"type": "text",
"bbox": [
107,
260,
504,
305
],
"lines": [
{
"bbox": [
105,
259,
506,
273
],
"spans": [
{
"bbox": [
105,
259,
327,
273
],
"score": 1.0,
"content": "One of the most popular neighbor embedding methods,",
"type": "text"
},
{
"bbox": [
327,
262,
333,
271
],
"score": 0.76,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
333,
259,
506,
273
],
"score": 1.0,
"content": "-distributed stochastic neighbor embedding",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
270,
506,
285
],
"spans": [
{
"bbox": [
105,
270,
110,
285
],
"score": 1.0,
"content": "(",
"type": "text"
},
{
"bbox": [
110,
272,
114,
282
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
115,
270,
210,
285
],
"score": 1.0,
"content": "-SNE; Van der Maaten",
"type": "text"
},
{
"bbox": [
211,
272,
221,
282
],
"score": 0.5,
"content": "\\&",
"type": "inline_equation"
},
{
"bbox": [
221,
270,
506,
285
],
"score": 1.0,
"content": "Hinton, 2008) is an extension of the stochastic neighbor embedding",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
283,
506,
295
],
"spans": [
{
"bbox": [
106,
283,
344,
295
],
"score": 1.0,
"content": "(SNE) originally suggested by Hinton & Roweis (2002).",
"type": "text"
},
{
"bbox": [
344,
283,
349,
293
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
349,
283,
506,
295
],
"score": 1.0,
"content": "-SNE minimizes the Kullback-Leibler",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
292,
446,
308
],
"spans": [
{
"bbox": [
105,
292,
397,
308
],
"score": 1.0,
"content": "divergence between the high-dimensional and low-dimensional affinities",
"type": "text"
},
{
"bbox": [
397,
295,
411,
306
],
"score": 0.87,
"content": "p _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
411,
292,
429,
308
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
429,
295,
441,
306
],
"score": 0.88,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
442,
292,
446,
308
],
"score": 1.0,
"content": ":",
"type": "text"
}
],
"index": 15
}
],
"index": 13.5,
"bbox_fs": [
105,
259,
506,
308
]
},
{
"type": "interline_equation",
"bbox": [
233,
311,
378,
339
],
"lines": [
{
"bbox": [
233,
311,
378,
339
],
"spans": [
{
"bbox": [
233,
311,
378,
339
],
"score": 0.95,
"content": "\\mathcal { L } = \\mathrm { { K L } } ( \\mathbf { P } \\parallel \\mathbf { Q } ) = \\sum _ { i j } p _ { i j } \\log \\frac { p _ { i j } } { q _ { i j } } .",
"type": "interline_equation",
"image_path": "84f384e8029d2b4f62221afe7a55436728d67f8a5b0edba25f989a97d956da3b.jpg"
}
]
}
],
"index": 16.5,
"virtual_lines": [
{
"bbox": [
233,
311,
378,
325.0
],
"spans": [],
"index": 16
},
{
"bbox": [
233,
325.0,
378,
339.0
],
"spans": [],
"index": 17
}
]
},
{
"type": "text",
"bbox": [
106,
344,
505,
378
],
"lines": [
{
"bbox": [
105,
344,
505,
357
],
"spans": [
{
"bbox": [
105,
344,
505,
357
],
"score": 1.0,
"content": "Both affinity matrices are defined to be symmetric, positive, and to sum to 1. In the original algo-",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
106,
356,
505,
368
],
"spans": [
{
"bbox": [
106,
356,
132,
368
],
"score": 1.0,
"content": "rithm,",
"type": "text"
},
{
"bbox": [
133,
356,
142,
366
],
"score": 0.62,
"content": "\\mathbf { P }",
"type": "inline_equation"
},
{
"bbox": [
142,
356,
505,
368
],
"score": 1.0,
"content": "was computed using adaptive Gaussian kernels, but almost the same results can be obtained",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
367,
500,
379
],
"spans": [
{
"bbox": [
105,
367,
288,
379
],
"score": 1.0,
"content": "simply by normalizing and symmetrizing the",
"type": "text"
},
{
"bbox": [
288,
367,
310,
377
],
"score": 0.74,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
310,
367,
500,
379
],
"score": 1.0,
"content": "graph adjacency matrix A (Bohm et al., 2022): ¨",
"type": "text"
}
],
"index": 20
}
],
"index": 19,
"bbox_fs": [
105,
344,
505,
379
]
},
{
"type": "interline_equation",
"bbox": [
262,
383,
349,
410
],
"lines": [
{
"bbox": [
262,
383,
349,
410
],
"spans": [
{
"bbox": [
262,
383,
349,
410
],
"score": 0.96,
"content": "\\mathbf { P } = { \\frac { \\mathbf { A } / k + \\mathbf { A } ^ { \\top } / k } { 2 n } } .",
"type": "interline_equation",
"image_path": "f689fa8500db6ec61882ece7b6b97aba88c5d29597b200782c6426f1484c3416.jpg"
}
]
}
],
"index": 21.5,
"virtual_lines": [
{
"bbox": [
262,
383,
349,
396.5
],
"spans": [],
"index": 21
},
{
"bbox": [
262,
396.5,
349,
410.0
],
"spans": [],
"index": 22
}
]
},
{
"type": "text",
"bbox": [
107,
414,
504,
447
],
"lines": [
{
"bbox": [
105,
414,
504,
427
],
"spans": [
{
"bbox": [
105,
414,
193,
427
],
"score": 1.0,
"content": "Here A has element",
"type": "text"
},
{
"bbox": [
194,
415,
229,
426
],
"score": 0.91,
"content": "a _ { i j } = 1",
"type": "inline_equation"
},
{
"bbox": [
229,
414,
240,
427
],
"score": 1.0,
"content": "if",
"type": "text"
},
{
"bbox": [
241,
415,
252,
426
],
"score": 0.86,
"content": "\\mathbf { x } _ { j }",
"type": "inline_equation"
},
{
"bbox": [
253,
414,
293,
427
],
"score": 1.0,
"content": "is within",
"type": "text"
},
{
"bbox": [
294,
415,
300,
424
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
301,
414,
389,
427
],
"score": 1.0,
"content": "nearest neighbors of",
"type": "text"
},
{
"bbox": [
389,
416,
399,
425
],
"score": 0.85,
"content": "\\mathbf { x } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
399,
414,
497,
427
],
"score": 1.0,
"content": ". Reasonable values of",
"type": "text"
},
{
"bbox": [
497,
415,
504,
424
],
"score": 0.83,
"content": "k",
"type": "inline_equation"
}
],
"index": 23
},
{
"bbox": [
105,
425,
506,
438
],
"spans": [
{
"bbox": [
105,
425,
361,
438
],
"score": 1.0,
"content": "typically lie between 10 and 100. Low-dimensional affinities",
"type": "text"
},
{
"bbox": [
361,
425,
372,
437
],
"score": 0.8,
"content": "\\mathbf { Q }",
"type": "inline_equation"
},
{
"bbox": [
372,
425,
433,
438
],
"score": 1.0,
"content": "are defined in",
"type": "text"
},
{
"bbox": [
433,
426,
438,
435
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
438,
425,
495,
438
],
"score": 1.0,
"content": "-SNE using a",
"type": "text"
},
{
"bbox": [
496,
426,
501,
435
],
"score": 0.73,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
501,
425,
506,
438
],
"score": 1.0,
"content": "-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
106,
436,
433,
448
],
"spans": [
{
"bbox": [
106,
436,
433,
448
],
"score": 1.0,
"content": "distribution kernel with one degree of freedom, also known as the Cauchy kernel:",
"type": "text"
}
],
"index": 25
}
],
"index": 24,
"bbox_fs": [
105,
414,
506,
448
]
},
{
"type": "interline_equation",
"bbox": [
236,
452,
375,
483
],
"lines": [
{
"bbox": [
236,
452,
375,
483
],
"spans": [
{
"bbox": [
236,
452,
375,
483
],
"score": 0.95,
"content": "q _ { i j } = \\frac { ( 1 + \\| \\mathbf { y } _ { i } - \\mathbf { y } _ { j } \\| ^ { 2 } ) ^ { - 1 } } { \\sum _ { k \\neq l } ( 1 + \\| \\mathbf { y } _ { l } - \\mathbf { y } _ { k } \\| ^ { 2 } ) ^ { - 1 } } .",
"type": "interline_equation",
"image_path": "c46e8a49680d153032d10a818f4f19b1812d9912d8c69d0757d68bfeb6234671.jpg"
}
]
}
],
"index": 26.5,
"virtual_lines": [
{
"bbox": [
236,
452,
375,
467.5
],
"spans": [],
"index": 26
},
{
"bbox": [
236,
467.5,
375,
483.0
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
487,
504,
510
],
"lines": [
{
"bbox": [
106,
487,
504,
500
],
"spans": [
{
"bbox": [
106,
487,
504,
500
],
"score": 1.0,
"content": "The original SNE algorithm used Gaussian kernel instead of Cauchy, which led to worse results",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
498,
353,
510
],
"spans": [
{
"bbox": [
106,
498,
353,
510
],
"score": 1.0,
"content": "when embedding high-dimensional data (Kobak et al., 2019).",
"type": "text"
}
],
"index": 29
}
],
"index": 28.5,
"bbox_fs": [
106,
487,
504,
510
]
},
{
"type": "text",
"bbox": [
106,
514,
505,
582
],
"lines": [
{
"bbox": [
106,
515,
505,
528
],
"spans": [
{
"bbox": [
106,
515,
297,
528
],
"score": 1.0,
"content": "Even though it is usually not presented like that,",
"type": "text"
},
{
"bbox": [
297,
516,
302,
525
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
303,
515,
505,
528
],
"score": 1.0,
"content": "-SNE can be thought of as a graph layout algorithm",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
526,
505,
539
],
"spans": [
{
"bbox": [
106,
526,
120,
539
],
"score": 1.0,
"content": "for",
"type": "text"
},
{
"bbox": [
121,
527,
142,
537
],
"score": 0.55,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
143,
526,
505,
539
],
"score": 1.0,
"content": "graphs, in particular after the reformulation in Equation 2. During optimization, neighbor-",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
537,
505,
550
],
"spans": [
{
"bbox": [
105,
537,
505,
550
],
"score": 1.0,
"content": "ing nodes (sharing an edge) feel attraction, whereas all nodes feel repulsion, arising through the",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
548,
505,
561
],
"spans": [
{
"bbox": [
105,
548,
273,
561
],
"score": 1.0,
"content": "normalization in Equation 3. In practice,",
"type": "text"
},
{
"bbox": [
273,
549,
279,
558
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
279,
548,
505,
561
],
"score": 1.0,
"content": "-SNE optimization can be accelerated by an approxima-",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
558,
506,
573
],
"spans": [
{
"bbox": [
105,
558,
506,
573
],
"score": 1.0,
"content": "tion of the repulsive force field based on the Barnes–Hut algorithm (Van Der Maaten, 2014; Yang",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
570,
335,
582
],
"spans": [
{
"bbox": [
106,
570,
335,
582
],
"score": 1.0,
"content": "et al., 2013) or on interpolation (Linderman et al., 2019).",
"type": "text"
}
],
"index": 35
}
],
"index": 32.5,
"bbox_fs": [
105,
515,
506,
582
]
},
{
"type": "title",
"bbox": [
109,
595,
302,
606
],
"lines": [
{
"bbox": [
106,
595,
302,
608
],
"spans": [
{
"bbox": [
106,
595,
302,
608
],
"score": 1.0,
"content": "3.2 CONTRASTIVE NEIGHBOR EMBEDDINGS",
"type": "text"
}
],
"index": 36
}
],
"index": 36
},
{
"type": "text",
"bbox": [
107,
615,
505,
704
],
"lines": [
{
"bbox": [
105,
615,
506,
629
],
"spans": [
{
"bbox": [
105,
615,
506,
629
],
"score": 1.0,
"content": "The contrastive neighbor embedding (CNE) algorithm (Damrich et al., 2022) is a flexible framework",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
627,
505,
639
],
"spans": [
{
"bbox": [
105,
627,
210,
639
],
"score": 1.0,
"content": "that also operates on the",
"type": "text"
},
{
"bbox": [
210,
627,
232,
637
],
"score": 0.34,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
233,
627,
505,
639
],
"score": 1.0,
"content": "graph of the data, and optimizes the embedding in order to place",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
637,
506,
650
],
"spans": [
{
"bbox": [
105,
637,
506,
650
],
"score": 1.0,
"content": "connected nodes closer together than unconnected pairs of nodes. Damrich et al. (2022) considered",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
648,
505,
661
],
"spans": [
{
"bbox": [
105,
648,
505,
661
],
"score": 1.0,
"content": "three different loss functions: NCE (noise-contrastive estimation) (Gutmann & Hyvarinen, 2010), ¨",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
659,
505,
673
],
"spans": [
{
"bbox": [
105,
659,
505,
673
],
"score": 1.0,
"content": "InfoNCE (Jozefowicz et al., 2016; Oord et al., 2018), and negative sampling (Mikolov et al., 2013).",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
670,
505,
684
],
"spans": [
{
"bbox": [
105,
670,
505,
684
],
"score": 1.0,
"content": "These loss functions are called contrastive because they are based on contrasting edges and non-",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
105,
680,
506,
695
],
"spans": [
{
"bbox": [
105,
680,
506,
695
],
"score": 1.0,
"content": "edges in the same mini-batch, and do not require a global normalization like in Equation 3. Using",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
692,
304,
705
],
"spans": [
{
"bbox": [
105,
692,
273,
705
],
"score": 1.0,
"content": "NCE and InfoNCE in CNE approximates",
"type": "text"
},
{
"bbox": [
274,
694,
278,
703
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
279,
692,
304,
705
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 44
}
],
"index": 40.5,
"bbox_fs": [
105,
615,
506,
705
]
},
{
"type": "text",
"bbox": [
107,
709,
504,
732
],
"lines": [
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "Damrich et al. (2022) also considered parametric embeddings, where a neural network (usually a",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
105,
720,
505,
734
],
"spans": [
{
"bbox": [
105,
720,
381,
734
],
"score": 1.0,
"content": "fully-connected network) is trained to produce embedding vectors",
"type": "text"
},
{
"bbox": [
381,
720,
433,
732
],
"score": 0.93,
"content": "\\mathbf { y } _ { i } ~ = ~ f ( \\mathbf { x } _ { i } )",
"type": "inline_equation"
},
{
"bbox": [
433,
720,
505,
734
],
"score": 1.0,
"content": "using one of the",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
226,
505,
237
],
"spans": [
{
"bbox": [
106,
226,
505,
237
],
"score": 1.0,
"content": "loss function listed above. This allows to embed new observations that have not been part of the",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
106,
236,
505,
249
],
"spans": [
{
"bbox": [
106,
236,
388,
249
],
"score": 1.0,
"content": "training process. In contrast, non-parametric embeddings optimize",
"type": "text",
"cross_page": true
},
{
"bbox": [
389,
238,
400,
248
],
"score": 0.84,
"content": "\\mathbf { y } _ { i }",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
401,
236,
505,
249
],
"score": 1.0,
"content": "vectors directly, without",
"type": "text",
"cross_page": true
}
],
"index": 6
},
{
"bbox": [
105,
247,
506,
260
],
"spans": [
{
"bbox": [
105,
247,
124,
260
],
"score": 1.0,
"content": "any",
"type": "text",
"cross_page": true
},
{
"bbox": [
124,
248,
142,
259
],
"score": 0.9,
"content": "f ( \\cdot )",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
142,
247,
506,
260
],
"score": 1.0,
"content": "function. Together, this yields six combinations, called parametric/non-parametric NC-",
"type": "text",
"cross_page": true
}
],
"index": 7
},
{
"bbox": [
106,
257,
505,
271
],
"spans": [
{
"bbox": [
106,
259,
111,
268
],
"score": 0.72,
"content": "t",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
112,
257,
171,
271
],
"score": 1.0,
"content": "-SNE, InfoNC-",
"type": "text",
"cross_page": true
},
{
"bbox": [
171,
259,
176,
268
],
"score": 0.61,
"content": "\\mathbf { \\nabla } \\cdot t",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
176,
257,
237,
271
],
"score": 1.0,
"content": "-SNE, and Neg-",
"type": "text",
"cross_page": true
},
{
"bbox": [
237,
259,
243,
268
],
"score": 0.42,
"content": "\\cdot t",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
243,
257,
424,
271
],
"score": 1.0,
"content": "-SNE. Damrich et al. (2022) showed that Neg-",
"type": "text",
"cross_page": true
},
{
"bbox": [
425,
259,
430,
268
],
"score": 0.48,
"content": "\\cdot t",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
430,
257,
505,
271
],
"score": 1.0,
"content": "-SNE is equivalent",
"type": "text",
"cross_page": true
}
],
"index": 8
},
{
"bbox": [
105,
269,
505,
282
],
"spans": [
{
"bbox": [
105,
269,
290,
282
],
"score": 1.0,
"content": "to UMAP (McInnes et al., 2018), while NC-",
"type": "text",
"cross_page": true
},
{
"bbox": [
290,
270,
295,
279
],
"score": 0.52,
"content": "\\cdot t .",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
296,
269,
505,
282
],
"score": 1.0,
"content": "-SNE was first suggested by Artemenkov & Panov",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
106,
280,
178,
292
],
"spans": [
{
"bbox": [
106,
280,
178,
292
],
"score": 1.0,
"content": "(2020) as NCVis.",
"type": "text",
"cross_page": true
}
],
"index": 10
}
],
"index": 45.5,
"bbox_fs": [
105,
709,
505,
734
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
106,
113,
505,
203
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
89,
502,
112
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
88,
504,
102
],
"spans": [
{
"bbox": [
106,
88,
504,
102
],
"score": 1.0,
"content": "Table 1: Benchmark datasets. Columns: number of nodes in the largest connected component,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
100,
489,
113
],
"spans": [
{
"bbox": [
105,
100,
489,
113
],
"score": 1.0,
"content": "number of undirected edges, edges/nodes ratio, number of node classes, feature dimensionality.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "table_body",
"bbox": [
106,
113,
505,
203
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
113,
505,
203
],
"spans": [
{
"bbox": [
106,
113,
505,
203
],
"score": 0.98,
"html": "
| Dataset | Abbr. | Nodes | Edges | E/N | Classes | Dim. |
| CiteseerGraphDataset | CSR | 2120 | 3679 | 1.7 | 6 | 3703 |
| CoraGraphDataset | COR | 2485 | 5069 | 2.0 | 7 | 1433 |
| AmazonCoBuyPhotoDataset | APH | 7487 | 119 043 | 15.9 | 8 | 745 |
| AmazonCoBuyComputerDataset | ACO | 13381 | 245 778 | 18.4 | 10 | 767 |
| PubmedGraphDataset | PUB | 19717 | 44324 | 2.2 | 3 | 500 |
| ogbn-arxiv | ARX | 169 343 | 1157799 | 6.8 | 40 | 128 |
",
"type": "table",
"image_path": "32045b7de92c000aa9741d404727a1fbdd02da3b666bb8192afd128d5335788f.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
106,
113,
505,
143.0
],
"spans": [],
"index": 2
},
{
"bbox": [
106,
143.0,
505,
173.0
],
"spans": [],
"index": 3
},
{
"bbox": [
106,
173.0,
505,
203.0
],
"spans": [],
"index": 4
}
]
}
],
"index": 1.75
},
{
"type": "text",
"bbox": [
106,
225,
505,
291
],
"lines": [
{
"bbox": [
106,
226,
505,
237
],
"spans": [
{
"bbox": [
106,
226,
505,
237
],
"score": 1.0,
"content": "loss function listed above. This allows to embed new observations that have not been part of the",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
106,
236,
505,
249
],
"spans": [
{
"bbox": [
106,
236,
388,
249
],
"score": 1.0,
"content": "training process. In contrast, non-parametric embeddings optimize",
"type": "text"
},
{
"bbox": [
389,
238,
400,
248
],
"score": 0.84,
"content": "\\mathbf { y } _ { i }",
"type": "inline_equation"
},
{
"bbox": [
401,
236,
505,
249
],
"score": 1.0,
"content": "vectors directly, without",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
247,
506,
260
],
"spans": [
{
"bbox": [
105,
247,
124,
260
],
"score": 1.0,
"content": "any",
"type": "text"
},
{
"bbox": [
124,
248,
142,
259
],
"score": 0.9,
"content": "f ( \\cdot )",
"type": "inline_equation"
},
{
"bbox": [
142,
247,
506,
260
],
"score": 1.0,
"content": "function. Together, this yields six combinations, called parametric/non-parametric NC-",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
257,
505,
271
],
"spans": [
{
"bbox": [
106,
259,
111,
268
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
112,
257,
171,
271
],
"score": 1.0,
"content": "-SNE, InfoNC-",
"type": "text"
},
{
"bbox": [
171,
259,
176,
268
],
"score": 0.61,
"content": "\\mathbf { \\nabla } \\cdot t",
"type": "inline_equation"
},
{
"bbox": [
176,
257,
237,
271
],
"score": 1.0,
"content": "-SNE, and Neg-",
"type": "text"
},
{
"bbox": [
237,
259,
243,
268
],
"score": 0.42,
"content": "\\cdot t",
"type": "inline_equation"
},
{
"bbox": [
243,
257,
424,
271
],
"score": 1.0,
"content": "-SNE. Damrich et al. (2022) showed that Neg-",
"type": "text"
},
{
"bbox": [
425,
259,
430,
268
],
"score": 0.48,
"content": "\\cdot t",
"type": "inline_equation"
},
{
"bbox": [
430,
257,
505,
271
],
"score": 1.0,
"content": "-SNE is equivalent",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
269,
505,
282
],
"spans": [
{
"bbox": [
105,
269,
290,
282
],
"score": 1.0,
"content": "to UMAP (McInnes et al., 2018), while NC-",
"type": "text"
},
{
"bbox": [
290,
270,
295,
279
],
"score": 0.52,
"content": "\\cdot t .",
"type": "inline_equation"
},
{
"bbox": [
296,
269,
505,
282
],
"score": 1.0,
"content": "-SNE was first suggested by Artemenkov & Panov",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
280,
178,
292
],
"spans": [
{
"bbox": [
106,
280,
178,
292
],
"score": 1.0,
"content": "(2020) as NCVis.",
"type": "text"
}
],
"index": 10
}
],
"index": 7.5
},
{
"type": "text",
"bbox": [
106,
297,
505,
319
],
"lines": [
{
"bbox": [
104,
295,
505,
310
],
"spans": [
{
"bbox": [
104,
295,
456,
310
],
"score": 1.0,
"content": "In this work we will only use the InfoNCE loss function, defined for one graph edge",
"type": "text"
},
{
"bbox": [
457,
298,
466,
309
],
"score": 0.72,
"content": "i j",
"type": "inline_equation"
},
{
"bbox": [
466,
295,
505,
310
],
"score": 1.0,
"content": "(positive",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
103,
308,
141,
321
],
"spans": [
{
"bbox": [
103,
308,
141,
321
],
"score": 1.0,
"content": "pair) as",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5
},
{
"type": "interline_equation",
"bbox": [
239,
320,
371,
345
],
"lines": [
{
"bbox": [
239,
320,
371,
345
],
"spans": [
{
"bbox": [
239,
320,
371,
345
],
"score": 0.95,
"content": "\\ell ( i , j ) = - \\log \\frac { q _ { i j } } { q _ { i j } + \\sum _ { k = 1 } ^ { m } q _ { i k } } ,",
"type": "interline_equation",
"image_path": "e7a10df242066fb864f77c049c0e98325858e908fda9983626b622d3c3ce927a.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
239,
320,
371,
345
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
106,
349,
505,
405
],
"lines": [
{
"bbox": [
105,
348,
505,
361
],
"spans": [
{
"bbox": [
105,
348,
273,
361
],
"score": 1.0,
"content": "where the sum in the denominator is over",
"type": "text"
},
{
"bbox": [
273,
352,
284,
359
],
"score": 0.55,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
284,
348,
343,
361
],
"score": 1.0,
"content": "negative pairs",
"type": "text"
},
{
"bbox": [
344,
350,
354,
360
],
"score": 0.85,
"content": "i k",
"type": "inline_equation"
},
{
"bbox": [
354,
348,
381,
361
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
382,
350,
389,
359
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
389,
348,
505,
361
],
"score": 1.0,
"content": "can be drawn from all nodes",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
359,
506,
373
],
"spans": [
{
"bbox": [
105,
359,
249,
373
],
"score": 1.0,
"content": "in the same mini-batch apart from",
"type": "text"
},
{
"bbox": [
249,
361,
254,
370
],
"score": 0.69,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
254,
359,
273,
373
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
273,
361,
280,
372
],
"score": 0.79,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
280,
359,
399,
373
],
"score": 1.0,
"content": ". One mini-batch consists of",
"type": "text"
},
{
"bbox": [
399,
361,
405,
370
],
"score": 0.71,
"content": "b",
"type": "inline_equation"
},
{
"bbox": [
406,
359,
506,
373
],
"score": 1.0,
"content": "graph edges, and hence",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
104,
369,
506,
385
],
"spans": [
{
"bbox": [
104,
369,
142,
385
],
"score": 1.0,
"content": "contains",
"type": "text"
},
{
"bbox": [
142,
371,
153,
381
],
"score": 0.55,
"content": "2 b",
"type": "inline_equation"
},
{
"bbox": [
154,
369,
314,
385
],
"score": 1.0,
"content": "nodes. Therefore, for a given batch size",
"type": "text"
},
{
"bbox": [
314,
372,
320,
381
],
"score": 0.75,
"content": "b",
"type": "inline_equation"
},
{
"bbox": [
320,
369,
409,
385
],
"score": 1.0,
"content": ", the maximal value of",
"type": "text"
},
{
"bbox": [
410,
372,
420,
381
],
"score": 0.78,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
420,
369,
430,
385
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
430,
371,
457,
382
],
"score": 0.85,
"content": "2 b - 2",
"type": "inline_equation"
},
{
"bbox": [
457,
369,
506,
385
],
"score": 1.0,
"content": ". The larger",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
381,
505,
395
],
"spans": [
{
"bbox": [
105,
381,
120,
395
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
121,
384,
131,
392
],
"score": 0.73,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
131,
381,
209,
395
],
"score": 1.0,
"content": ", the closer InfoNC-",
"type": "text"
},
{
"bbox": [
209,
383,
214,
392
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
214,
381,
256,
395
],
"score": 1.0,
"content": "-SNE is to",
"type": "text"
},
{
"bbox": [
256,
383,
261,
392
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
262,
381,
395,
395
],
"score": 1.0,
"content": "-SNE (Damrich et al., 2022). The",
"type": "text"
},
{
"bbox": [
396,
383,
408,
394
],
"score": 0.86,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
408,
381,
505,
395
],
"score": 1.0,
"content": "affinities do not need to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
392,
270,
406
],
"spans": [
{
"bbox": [
105,
392,
270,
406
],
"score": 1.0,
"content": "be normalized and are defined simply as",
"type": "text"
}
],
"index": 18
}
],
"index": 16
},
{
"type": "interline_equation",
"bbox": [
249,
410,
362,
426
],
"lines": [
{
"bbox": [
249,
410,
362,
426
],
"spans": [
{
"bbox": [
249,
410,
362,
426
],
"score": 0.92,
"content": "q _ { i j } = ( 1 + \\| \\mathbf { y } _ { i } - \\mathbf { y } _ { j } \\| ^ { 2 } ) ^ { - 1 } .",
"type": "interline_equation",
"image_path": "542f3326e12a976e1b1d73e77d6ac81f195d71cc97e9aaf70bc2c7a823f0a235.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
249,
410,
362,
426
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
107,
432,
505,
455
],
"lines": [
{
"bbox": [
105,
432,
505,
445
],
"spans": [
{
"bbox": [
105,
432,
314,
444
],
"score": 1.0,
"content": "It is easy to see that InfoNCE loss will aim to make",
"type": "text"
},
{
"bbox": [
315,
434,
327,
445
],
"score": 0.88,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
327,
432,
360,
444
],
"score": 1.0,
"content": "large if",
"type": "text"
},
{
"bbox": [
360,
433,
369,
444
],
"score": 0.88,
"content": "i j",
"type": "inline_equation"
},
{
"bbox": [
369,
432,
505,
444
],
"score": 1.0,
"content": "is a positive pair and small if it is",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
443,
168,
455
],
"spans": [
{
"bbox": [
105,
443,
168,
455
],
"score": 1.0,
"content": "a negative one.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5
},
{
"type": "text",
"bbox": [
107,
460,
505,
493
],
"lines": [
{
"bbox": [
106,
460,
505,
472
],
"spans": [
{
"bbox": [
106,
460,
327,
472
],
"score": 1.0,
"content": "When using high-dimensional embedding space, e.g.",
"type": "text"
},
{
"bbox": [
327,
460,
366,
470
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
366,
460,
411,
472
],
"score": 1.0,
"content": "instead of",
"type": "text"
},
{
"bbox": [
411,
460,
438,
470
],
"score": 0.9,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
439,
460,
505,
472
],
"score": 1.0,
"content": ", it makes sense",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
144,
484
],
"score": 1.0,
"content": "to define",
"type": "text"
},
{
"bbox": [
144,
473,
156,
483
],
"score": 0.87,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
157,
471,
505,
484
],
"score": 1.0,
"content": "using the Gaussian kernel transformation of the cosine distance (Damrich et al., 2022;",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
482,
186,
493
],
"spans": [
{
"bbox": [
106,
482,
186,
493
],
"score": 1.0,
"content": "Bohm et al., 2023):¨",
"type": "text"
}
],
"index": 24
}
],
"index": 23
},
{
"type": "interline_equation",
"bbox": [
141,
500,
469,
527
],
"lines": [
{
"bbox": [
141,
500,
469,
527
],
"spans": [
{
"bbox": [
141,
500,
469,
527
],
"score": 0.92,
"content": "q _ { i j } = \\exp \\bigl ( \\mathbf { y } _ { i } ^ { \\mathsf { T } } \\mathbf { y } _ { j } / ( \\lVert \\mathbf { y } _ { i } \\rVert \\cdot \\lVert \\mathbf { y } _ { j } \\rVert ) / \\tau \\bigr ) = \\mathrm { c o n s t } \\cdot \\exp \\Big ( - \\Big \\lVert \\frac { \\mathbf { y } _ { i } } { \\lVert \\mathbf { y } _ { i } \\rVert } - \\frac { \\mathbf { y } _ { j } } { \\lVert \\mathbf { y } _ { j } \\rVert } \\Big \\rVert ^ { 2 } \\Big / ( 2 \\tau ) \\Big ) ,",
"type": "interline_equation",
"image_path": "f1aeccf40854bc160dcf59105db1a796b87b92147e9ccbcdc96975959727fbc8.jpg"
}
]
}
],
"index": 26,
"virtual_lines": [
{
"bbox": [
141,
500,
469,
509.0
],
"spans": [],
"index": 25
},
{
"bbox": [
141,
509.0,
469,
518.0
],
"spans": [],
"index": 26
},
{
"bbox": [
141,
518.0,
469,
527.0
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
533,
505,
578
],
"lines": [
{
"bbox": [
106,
533,
505,
546
],
"spans": [
{
"bbox": [
106,
533,
133,
546
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
536,
141,
543
],
"score": 0.77,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
141,
533,
295,
546
],
"score": 1.0,
"content": "is called the temperature (by default,",
"type": "text"
},
{
"bbox": [
295,
533,
331,
544
],
"score": 0.87,
"content": "\\tau = 0 . 5",
"type": "inline_equation"
},
{
"bbox": [
331,
533,
505,
546
],
"score": 1.0,
"content": "). Together with Equation 5, this gives the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
543,
505,
557
],
"spans": [
{
"bbox": [
105,
543,
505,
557
],
"score": 1.0,
"content": "same loss function as used in SimCLR (Chen et al., 2020), a popular contrastive learning algo-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
555,
506,
568
],
"spans": [
{
"bbox": [
106,
556,
363,
568
],
"score": 1.0,
"content": "rithm in computer vision. The only difference is that instead of",
"type": "text"
},
{
"bbox": [
363,
555,
385,
566
],
"score": 0.78,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
385,
556,
506,
568
],
"score": 1.0,
"content": "edges, SimCLR uses pairs of",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
567,
252,
579
],
"spans": [
{
"bbox": [
105,
567,
252,
579
],
"score": 1.0,
"content": "augmented images as positive pairs.",
"type": "text"
}
],
"index": 31
}
],
"index": 29.5
},
{
"type": "title",
"bbox": [
108,
595,
244,
608
],
"lines": [
{
"bbox": [
105,
594,
245,
610
],
"spans": [
{
"bbox": [
105,
594,
245,
610
],
"score": 1.0,
"content": "4 EXPERIMENTAL SETUP",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "text",
"bbox": [
106,
621,
505,
732
],
"lines": [
{
"bbox": [
105,
621,
505,
634
],
"spans": [
{
"bbox": [
105,
621,
505,
634
],
"score": 1.0,
"content": "Datasets We used six publicly available graph datasets (Table 1). All datasets were retrieved from",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
633,
505,
646
],
"spans": [
{
"bbox": [
105,
633,
505,
646
],
"score": 1.0,
"content": "the Deep Graph Library (Wang et al., 2019), except ogbn-arxiv, which was retrieved from the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
643,
506,
656
],
"spans": [
{
"bbox": [
106,
643,
506,
656
],
"score": 1.0,
"content": "Open Graph Benchmark (Hu et al., 2020). Each dataset was treated as an unweighted undirected",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
654,
506,
668
],
"spans": [
{
"bbox": [
105,
654,
506,
668
],
"score": 1.0,
"content": "graph, where each node has a class label and a feature vector (typically a word embedding vector",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
666,
505,
678
],
"spans": [
{
"bbox": [
106,
666,
505,
678
],
"score": 1.0,
"content": "of some descriptive text about the node, such as a product review). We restricted ourselves to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
677,
506,
690
],
"spans": [
{
"bbox": [
105,
677,
506,
690
],
"score": 1.0,
"content": "graphs with labeled nodes in order to use classification accuracy as the performance metric. We also",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "restricted ourselves to graphs with feature vectors in order to use both non-parametric and parametric",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
699,
505,
712
],
"spans": [
{
"bbox": [
105,
699,
505,
712
],
"score": 1.0,
"content": "embeddings. In all datasets we used only the largest connected component, and excluded all self-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
710,
506,
722
],
"spans": [
{
"bbox": [
105,
710,
506,
722
],
"score": 1.0,
"content": "loops if present, using NetworkX (Hagberg et al., 2008) functions connected components",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
720,
209,
733
],
"spans": [
{
"bbox": [
106,
720,
209,
733
],
"score": 1.0,
"content": "and selfloop edges.",
"type": "text"
}
],
"index": 42
}
],
"index": 37.5
}
],
"page_idx": 3,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
750,
310,
762
],
"spans": [
{
"bbox": [
302,
750,
310,
762
],
"score": 1.0,
"content": "",
"type": "text",
"height": 12,
"width": 8
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
106,
113,
505,
203
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
89,
502,
112
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
88,
504,
102
],
"spans": [
{
"bbox": [
106,
88,
504,
102
],
"score": 1.0,
"content": "Table 1: Benchmark datasets. Columns: number of nodes in the largest connected component,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
105,
100,
489,
113
],
"spans": [
{
"bbox": [
105,
100,
489,
113
],
"score": 1.0,
"content": "number of undirected edges, edges/nodes ratio, number of node classes, feature dimensionality.",
"type": "text"
}
],
"index": 1
}
],
"index": 0.5
},
{
"type": "table_body",
"bbox": [
106,
113,
505,
203
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
113,
505,
203
],
"spans": [
{
"bbox": [
106,
113,
505,
203
],
"score": 0.98,
"html": "| Dataset | Abbr. | Nodes | Edges | E/N | Classes | Dim. |
| CiteseerGraphDataset | CSR | 2120 | 3679 | 1.7 | 6 | 3703 |
| CoraGraphDataset | COR | 2485 | 5069 | 2.0 | 7 | 1433 |
| AmazonCoBuyPhotoDataset | APH | 7487 | 119 043 | 15.9 | 8 | 745 |
| AmazonCoBuyComputerDataset | ACO | 13381 | 245 778 | 18.4 | 10 | 767 |
| PubmedGraphDataset | PUB | 19717 | 44324 | 2.2 | 3 | 500 |
| ogbn-arxiv | ARX | 169 343 | 1157799 | 6.8 | 40 | 128 |
",
"type": "table",
"image_path": "32045b7de92c000aa9741d404727a1fbdd02da3b666bb8192afd128d5335788f.jpg"
}
]
}
],
"index": 3,
"virtual_lines": [
{
"bbox": [
106,
113,
505,
143.0
],
"spans": [],
"index": 2
},
{
"bbox": [
106,
143.0,
505,
173.0
],
"spans": [],
"index": 3
},
{
"bbox": [
106,
173.0,
505,
203.0
],
"spans": [],
"index": 4
}
]
}
],
"index": 1.75
},
{
"type": "text",
"bbox": [
106,
225,
505,
291
],
"lines": [],
"index": 7.5,
"bbox_fs": [
105,
226,
506,
292
],
"lines_deleted": true
},
{
"type": "text",
"bbox": [
106,
297,
505,
319
],
"lines": [
{
"bbox": [
104,
295,
505,
310
],
"spans": [
{
"bbox": [
104,
295,
456,
310
],
"score": 1.0,
"content": "In this work we will only use the InfoNCE loss function, defined for one graph edge",
"type": "text"
},
{
"bbox": [
457,
298,
466,
309
],
"score": 0.72,
"content": "i j",
"type": "inline_equation"
},
{
"bbox": [
466,
295,
505,
310
],
"score": 1.0,
"content": "(positive",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
103,
308,
141,
321
],
"spans": [
{
"bbox": [
103,
308,
141,
321
],
"score": 1.0,
"content": "pair) as",
"type": "text"
}
],
"index": 12
}
],
"index": 11.5,
"bbox_fs": [
103,
295,
505,
321
]
},
{
"type": "interline_equation",
"bbox": [
239,
320,
371,
345
],
"lines": [
{
"bbox": [
239,
320,
371,
345
],
"spans": [
{
"bbox": [
239,
320,
371,
345
],
"score": 0.95,
"content": "\\ell ( i , j ) = - \\log \\frac { q _ { i j } } { q _ { i j } + \\sum _ { k = 1 } ^ { m } q _ { i k } } ,",
"type": "interline_equation",
"image_path": "e7a10df242066fb864f77c049c0e98325858e908fda9983626b622d3c3ce927a.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
239,
320,
371,
345
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
106,
349,
505,
405
],
"lines": [
{
"bbox": [
105,
348,
505,
361
],
"spans": [
{
"bbox": [
105,
348,
273,
361
],
"score": 1.0,
"content": "where the sum in the denominator is over",
"type": "text"
},
{
"bbox": [
273,
352,
284,
359
],
"score": 0.55,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
284,
348,
343,
361
],
"score": 1.0,
"content": "negative pairs",
"type": "text"
},
{
"bbox": [
344,
350,
354,
360
],
"score": 0.85,
"content": "i k",
"type": "inline_equation"
},
{
"bbox": [
354,
348,
381,
361
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
382,
350,
389,
359
],
"score": 0.82,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
389,
348,
505,
361
],
"score": 1.0,
"content": "can be drawn from all nodes",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
359,
506,
373
],
"spans": [
{
"bbox": [
105,
359,
249,
373
],
"score": 1.0,
"content": "in the same mini-batch apart from",
"type": "text"
},
{
"bbox": [
249,
361,
254,
370
],
"score": 0.69,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
254,
359,
273,
373
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
273,
361,
280,
372
],
"score": 0.79,
"content": "j",
"type": "inline_equation"
},
{
"bbox": [
280,
359,
399,
373
],
"score": 1.0,
"content": ". One mini-batch consists of",
"type": "text"
},
{
"bbox": [
399,
361,
405,
370
],
"score": 0.71,
"content": "b",
"type": "inline_equation"
},
{
"bbox": [
406,
359,
506,
373
],
"score": 1.0,
"content": "graph edges, and hence",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
104,
369,
506,
385
],
"spans": [
{
"bbox": [
104,
369,
142,
385
],
"score": 1.0,
"content": "contains",
"type": "text"
},
{
"bbox": [
142,
371,
153,
381
],
"score": 0.55,
"content": "2 b",
"type": "inline_equation"
},
{
"bbox": [
154,
369,
314,
385
],
"score": 1.0,
"content": "nodes. Therefore, for a given batch size",
"type": "text"
},
{
"bbox": [
314,
372,
320,
381
],
"score": 0.75,
"content": "b",
"type": "inline_equation"
},
{
"bbox": [
320,
369,
409,
385
],
"score": 1.0,
"content": ", the maximal value of",
"type": "text"
},
{
"bbox": [
410,
372,
420,
381
],
"score": 0.78,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
420,
369,
430,
385
],
"score": 1.0,
"content": "is",
"type": "text"
},
{
"bbox": [
430,
371,
457,
382
],
"score": 0.85,
"content": "2 b - 2",
"type": "inline_equation"
},
{
"bbox": [
457,
369,
506,
385
],
"score": 1.0,
"content": ". The larger",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
381,
505,
395
],
"spans": [
{
"bbox": [
105,
381,
120,
395
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
121,
384,
131,
392
],
"score": 0.73,
"content": "m",
"type": "inline_equation"
},
{
"bbox": [
131,
381,
209,
395
],
"score": 1.0,
"content": ", the closer InfoNC-",
"type": "text"
},
{
"bbox": [
209,
383,
214,
392
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
214,
381,
256,
395
],
"score": 1.0,
"content": "-SNE is to",
"type": "text"
},
{
"bbox": [
256,
383,
261,
392
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
262,
381,
395,
395
],
"score": 1.0,
"content": "-SNE (Damrich et al., 2022). The",
"type": "text"
},
{
"bbox": [
396,
383,
408,
394
],
"score": 0.86,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
408,
381,
505,
395
],
"score": 1.0,
"content": "affinities do not need to",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
392,
270,
406
],
"spans": [
{
"bbox": [
105,
392,
270,
406
],
"score": 1.0,
"content": "be normalized and are defined simply as",
"type": "text"
}
],
"index": 18
}
],
"index": 16,
"bbox_fs": [
104,
348,
506,
406
]
},
{
"type": "interline_equation",
"bbox": [
249,
410,
362,
426
],
"lines": [
{
"bbox": [
249,
410,
362,
426
],
"spans": [
{
"bbox": [
249,
410,
362,
426
],
"score": 0.92,
"content": "q _ { i j } = ( 1 + \\| \\mathbf { y } _ { i } - \\mathbf { y } _ { j } \\| ^ { 2 } ) ^ { - 1 } .",
"type": "interline_equation",
"image_path": "542f3326e12a976e1b1d73e77d6ac81f195d71cc97e9aaf70bc2c7a823f0a235.jpg"
}
]
}
],
"index": 19,
"virtual_lines": [
{
"bbox": [
249,
410,
362,
426
],
"spans": [],
"index": 19
}
]
},
{
"type": "text",
"bbox": [
107,
432,
505,
455
],
"lines": [
{
"bbox": [
105,
432,
505,
445
],
"spans": [
{
"bbox": [
105,
432,
314,
444
],
"score": 1.0,
"content": "It is easy to see that InfoNCE loss will aim to make",
"type": "text"
},
{
"bbox": [
315,
434,
327,
445
],
"score": 0.88,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
327,
432,
360,
444
],
"score": 1.0,
"content": "large if",
"type": "text"
},
{
"bbox": [
360,
433,
369,
444
],
"score": 0.88,
"content": "i j",
"type": "inline_equation"
},
{
"bbox": [
369,
432,
505,
444
],
"score": 1.0,
"content": "is a positive pair and small if it is",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
443,
168,
455
],
"spans": [
{
"bbox": [
105,
443,
168,
455
],
"score": 1.0,
"content": "a negative one.",
"type": "text"
}
],
"index": 21
}
],
"index": 20.5,
"bbox_fs": [
105,
432,
505,
455
]
},
{
"type": "text",
"bbox": [
107,
460,
505,
493
],
"lines": [
{
"bbox": [
106,
460,
505,
472
],
"spans": [
{
"bbox": [
106,
460,
327,
472
],
"score": 1.0,
"content": "When using high-dimensional embedding space, e.g.",
"type": "text"
},
{
"bbox": [
327,
460,
366,
470
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
366,
460,
411,
472
],
"score": 1.0,
"content": "instead of",
"type": "text"
},
{
"bbox": [
411,
460,
438,
470
],
"score": 0.9,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
439,
460,
505,
472
],
"score": 1.0,
"content": ", it makes sense",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
471,
505,
484
],
"spans": [
{
"bbox": [
105,
471,
144,
484
],
"score": 1.0,
"content": "to define",
"type": "text"
},
{
"bbox": [
144,
473,
156,
483
],
"score": 0.87,
"content": "q _ { i j }",
"type": "inline_equation"
},
{
"bbox": [
157,
471,
505,
484
],
"score": 1.0,
"content": "using the Gaussian kernel transformation of the cosine distance (Damrich et al., 2022;",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
482,
186,
493
],
"spans": [
{
"bbox": [
106,
482,
186,
493
],
"score": 1.0,
"content": "Bohm et al., 2023):¨",
"type": "text"
}
],
"index": 24
}
],
"index": 23,
"bbox_fs": [
105,
460,
505,
493
]
},
{
"type": "interline_equation",
"bbox": [
141,
500,
469,
527
],
"lines": [
{
"bbox": [
141,
500,
469,
527
],
"spans": [
{
"bbox": [
141,
500,
469,
527
],
"score": 0.92,
"content": "q _ { i j } = \\exp \\bigl ( \\mathbf { y } _ { i } ^ { \\mathsf { T } } \\mathbf { y } _ { j } / ( \\lVert \\mathbf { y } _ { i } \\rVert \\cdot \\lVert \\mathbf { y } _ { j } \\rVert ) / \\tau \\bigr ) = \\mathrm { c o n s t } \\cdot \\exp \\Big ( - \\Big \\lVert \\frac { \\mathbf { y } _ { i } } { \\lVert \\mathbf { y } _ { i } \\rVert } - \\frac { \\mathbf { y } _ { j } } { \\lVert \\mathbf { y } _ { j } \\rVert } \\Big \\rVert ^ { 2 } \\Big / ( 2 \\tau ) \\Big ) ,",
"type": "interline_equation",
"image_path": "f1aeccf40854bc160dcf59105db1a796b87b92147e9ccbcdc96975959727fbc8.jpg"
}
]
}
],
"index": 26,
"virtual_lines": [
{
"bbox": [
141,
500,
469,
509.0
],
"spans": [],
"index": 25
},
{
"bbox": [
141,
509.0,
469,
518.0
],
"spans": [],
"index": 26
},
{
"bbox": [
141,
518.0,
469,
527.0
],
"spans": [],
"index": 27
}
]
},
{
"type": "text",
"bbox": [
106,
533,
505,
578
],
"lines": [
{
"bbox": [
106,
533,
505,
546
],
"spans": [
{
"bbox": [
106,
533,
133,
546
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
536,
141,
543
],
"score": 0.77,
"content": "\\tau",
"type": "inline_equation"
},
{
"bbox": [
141,
533,
295,
546
],
"score": 1.0,
"content": "is called the temperature (by default,",
"type": "text"
},
{
"bbox": [
295,
533,
331,
544
],
"score": 0.87,
"content": "\\tau = 0 . 5",
"type": "inline_equation"
},
{
"bbox": [
331,
533,
505,
546
],
"score": 1.0,
"content": "). Together with Equation 5, this gives the",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
543,
505,
557
],
"spans": [
{
"bbox": [
105,
543,
505,
557
],
"score": 1.0,
"content": "same loss function as used in SimCLR (Chen et al., 2020), a popular contrastive learning algo-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
555,
506,
568
],
"spans": [
{
"bbox": [
106,
556,
363,
568
],
"score": 1.0,
"content": "rithm in computer vision. The only difference is that instead of",
"type": "text"
},
{
"bbox": [
363,
555,
385,
566
],
"score": 0.78,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
385,
556,
506,
568
],
"score": 1.0,
"content": "edges, SimCLR uses pairs of",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
567,
252,
579
],
"spans": [
{
"bbox": [
105,
567,
252,
579
],
"score": 1.0,
"content": "augmented images as positive pairs.",
"type": "text"
}
],
"index": 31
}
],
"index": 29.5,
"bbox_fs": [
105,
533,
506,
579
]
},
{
"type": "title",
"bbox": [
108,
595,
244,
608
],
"lines": [
{
"bbox": [
105,
594,
245,
610
],
"spans": [
{
"bbox": [
105,
594,
245,
610
],
"score": 1.0,
"content": "4 EXPERIMENTAL SETUP",
"type": "text"
}
],
"index": 32
}
],
"index": 32
},
{
"type": "text",
"bbox": [
106,
621,
505,
732
],
"lines": [
{
"bbox": [
105,
621,
505,
634
],
"spans": [
{
"bbox": [
105,
621,
505,
634
],
"score": 1.0,
"content": "Datasets We used six publicly available graph datasets (Table 1). All datasets were retrieved from",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
633,
505,
646
],
"spans": [
{
"bbox": [
105,
633,
505,
646
],
"score": 1.0,
"content": "the Deep Graph Library (Wang et al., 2019), except ogbn-arxiv, which was retrieved from the",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
643,
506,
656
],
"spans": [
{
"bbox": [
106,
643,
506,
656
],
"score": 1.0,
"content": "Open Graph Benchmark (Hu et al., 2020). Each dataset was treated as an unweighted undirected",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
654,
506,
668
],
"spans": [
{
"bbox": [
105,
654,
506,
668
],
"score": 1.0,
"content": "graph, where each node has a class label and a feature vector (typically a word embedding vector",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
106,
666,
505,
678
],
"spans": [
{
"bbox": [
106,
666,
505,
678
],
"score": 1.0,
"content": "of some descriptive text about the node, such as a product review). We restricted ourselves to",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
677,
506,
690
],
"spans": [
{
"bbox": [
105,
677,
506,
690
],
"score": 1.0,
"content": "graphs with labeled nodes in order to use classification accuracy as the performance metric. We also",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
687,
505,
700
],
"spans": [
{
"bbox": [
105,
687,
505,
700
],
"score": 1.0,
"content": "restricted ourselves to graphs with feature vectors in order to use both non-parametric and parametric",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
699,
505,
712
],
"spans": [
{
"bbox": [
105,
699,
505,
712
],
"score": 1.0,
"content": "embeddings. In all datasets we used only the largest connected component, and excluded all self-",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
710,
506,
722
],
"spans": [
{
"bbox": [
105,
710,
506,
722
],
"score": 1.0,
"content": "loops if present, using NetworkX (Hagberg et al., 2008) functions connected components",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
106,
720,
209,
733
],
"spans": [
{
"bbox": [
106,
720,
209,
733
],
"score": 1.0,
"content": "and selfloop edges.",
"type": "text"
}
],
"index": 42
}
],
"index": 37.5,
"bbox_fs": [
105,
621,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
104,
75,
496,
294
],
"blocks": [
{
"type": "image_body",
"bbox": [
104,
75,
496,
294
],
"group_id": 0,
"lines": [
{
"bbox": [
104,
75,
496,
294
],
"spans": [
{
"bbox": [
104,
75,
496,
294
],
"score": 0.973,
"type": "image",
"image_path": "2e5f7f25dc75ee881bccb233b1ff2c656b23adc633b4fd11cb1bd5536ccf62fd.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
104,
75,
496,
148.0
],
"spans": [],
"index": 0
},
{
"bbox": [
104,
148.0,
496,
221.0
],
"spans": [],
"index": 1
},
{
"bbox": [
104,
221.0,
496,
294.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
314,
505,
348
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
314,
504,
326
],
"spans": [
{
"bbox": [
106,
314,
504,
326
],
"score": 1.0,
"content": "Figure 2: Embeddings of the ACO and APH datasets obtained using FDP (Fruchterman & Rein-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
325,
505,
338
],
"spans": [
{
"bbox": [
105,
325,
299,
338
],
"score": 1.0,
"content": "gold, 1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
299,
326,
304,
336
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
305,
325,
474,
338
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), and our graph",
"type": "text"
},
{
"bbox": [
474,
326,
479,
335
],
"score": 0.58,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
480,
325,
505,
338
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
336,
505,
349
],
"spans": [
{
"bbox": [
106,
336,
505,
349
],
"score": 1.0,
"content": "Embeddings in each row were aligned using Procrustes rotation. See Figure A.3 for all six datasets.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
106,
366,
505,
400
],
"lines": [
{
"bbox": [
105,
365,
505,
380
],
"spans": [
{
"bbox": [
105,
365,
505,
380
],
"score": 1.0,
"content": "Performance metrics We evaluated the performance of our methods using three performance",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
378,
505,
391
],
"spans": [
{
"bbox": [
105,
378,
142,
391
],
"score": 1.0,
"content": "metrics:",
"type": "text"
},
{
"bbox": [
142,
379,
149,
388
],
"score": 0.79,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
149,
378,
226,
391
],
"score": 1.0,
"content": "-nearest-neighbors",
"type": "text"
},
{
"bbox": [
226,
378,
252,
389
],
"score": 0.32,
"content": "( k \\mathsf { N N } )",
"type": "inline_equation"
},
{
"bbox": [
253,
378,
281,
391
],
"score": 1.0,
"content": "recall,",
"type": "text"
},
{
"bbox": [
281,
378,
303,
388
],
"score": 0.54,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
303,
378,
505,
391
],
"score": 1.0,
"content": "classification accuracy, and, for high-dimensional",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
388,
279,
402
],
"spans": [
{
"bbox": [
106,
388,
279,
402
],
"score": 1.0,
"content": "embeddings, linear classification accuracy.",
"type": "text"
}
],
"index": 8
}
],
"index": 7
},
{
"type": "text",
"bbox": [
107,
405,
505,
439
],
"lines": [
{
"bbox": [
106,
406,
504,
418
],
"spans": [
{
"bbox": [
106,
406,
124,
418
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
124,
406,
146,
416
],
"score": 0.69,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
406,
504,
418
],
"score": 1.0,
"content": "recall quantifies how well local node neighborhoods are preserved in the embedding. We",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
416,
505,
429
],
"spans": [
{
"bbox": [
105,
416,
505,
429
],
"score": 1.0,
"content": "defined it as the average fraction of each node’s graph neighbors that are among the node’s nearest",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
428,
222,
441
],
"spans": [
{
"bbox": [
106,
428,
222,
441
],
"score": 1.0,
"content": "neighbors in the embedding:",
"type": "text"
}
],
"index": 11
}
],
"index": 10
},
{
"type": "interline_equation",
"bbox": [
227,
443,
382,
478
],
"lines": [
{
"bbox": [
227,
443,
382,
478
],
"spans": [
{
"bbox": [
227,
443,
382,
478
],
"score": 0.95,
"content": "\\mathrm { R e c a l l } = \\frac { 1 } { | \\mathcal { V } | } \\sum _ { i = 1 } ^ { | \\mathcal { V } | } \\frac { \\left| N _ { G } [ i ] \\cap N _ { E , k _ { i } } [ i ] \\right| } { k _ { i } } ,",
"type": "interline_equation",
"image_path": "013acbeb5f8e7077d62b07cab43f72279519c94d2eac7331ea5a660e788ee069.jpg"
}
]
}
],
"index": 12.5,
"virtual_lines": [
{
"bbox": [
227,
443,
382,
460.5
],
"spans": [],
"index": 12
},
{
"bbox": [
227,
460.5,
382,
478.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
107,
481,
505,
538
],
"lines": [
{
"bbox": [
105,
480,
504,
495
],
"spans": [
{
"bbox": [
105,
480,
133,
495
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
482,
147,
494
],
"score": 0.9,
"content": "| \\nu |",
"type": "inline_equation"
},
{
"bbox": [
147,
480,
294,
495
],
"score": 1.0,
"content": "is the number of nodes in the graph,",
"type": "text"
},
{
"bbox": [
294,
482,
319,
494
],
"score": 0.92,
"content": "N _ { G } [ i ]",
"type": "inline_equation"
},
{
"bbox": [
319,
480,
390,
495
],
"score": 1.0,
"content": "is the set of node",
"type": "text"
},
{
"bbox": [
390,
483,
395,
492
],
"score": 0.64,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
395,
480,
472,
495
],
"score": 1.0,
"content": "’s graph neighbors,",
"type": "text"
},
{
"bbox": [
472,
482,
504,
493
],
"score": 0.89,
"content": "N _ { E , k } [ i ]",
"type": "inline_equation"
}
],
"index": 14
},
{
"bbox": [
105,
490,
504,
507
],
"spans": [
{
"bbox": [
105,
490,
200,
507
],
"score": 1.0,
"content": "denotes the set of node",
"type": "text"
},
{
"bbox": [
200,
494,
205,
503
],
"score": 0.71,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
205,
490,
212,
507
],
"score": 1.0,
"content": "’s",
"type": "text"
},
{
"bbox": [
213,
493,
220,
503
],
"score": 0.8,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
220,
490,
451,
507
],
"score": 1.0,
"content": "Euclidean nearest neighbors in the embedding space, and",
"type": "text"
},
{
"bbox": [
452,
493,
504,
505
],
"score": 0.9,
"content": "k _ { i } = | N _ { G } [ i ] |",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
105,
504,
505,
516
],
"spans": [
{
"bbox": [
105,
504,
197,
516
],
"score": 1.0,
"content": "is the number of node",
"type": "text"
},
{
"bbox": [
198,
505,
203,
514
],
"score": 0.71,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
203,
504,
505,
516
],
"score": 1.0,
"content": "’s graph neighbors. This metric is similar to what is commonly used in the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
514,
506,
528
],
"spans": [
{
"bbox": [
105,
514,
506,
528
],
"score": 1.0,
"content": "literature to benchmark graph layout algorithms (Kruiger et al., 2017; Zhu et al., 2020a; Zhong et al.,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
525,
380,
539
],
"spans": [
{
"bbox": [
106,
525,
380,
539
],
"score": 1.0,
"content": "2023), and so is our main metric for measuring graph layout quality.",
"type": "text"
}
],
"index": 18
}
],
"index": 16
},
{
"type": "text",
"bbox": [
107,
542,
505,
609
],
"lines": [
{
"bbox": [
106,
542,
505,
555
],
"spans": [
{
"bbox": [
106,
542,
125,
555
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
125,
543,
147,
553
],
"score": 0.31,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
542,
505,
555
],
"score": 1.0,
"content": "classification accuracy quantifies local class separation in the embedding. To calculate",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
107,
553,
505,
566
],
"spans": [
{
"bbox": [
107,
554,
128,
564
],
"score": 0.51,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
128,
553,
297,
566
],
"score": 1.0,
"content": "accuracy, we split all nodes into a training",
"type": "text"
},
{
"bbox": [
298,
554,
312,
566
],
"score": 0.8,
"content": "( 2 / 3",
"type": "inline_equation"
},
{
"bbox": [
313,
553,
417,
566
],
"score": 1.0,
"content": "of all nodes) and a test set",
"type": "text"
},
{
"bbox": [
418,
554,
433,
565
],
"score": 0.83,
"content": "( 1 / 3",
"type": "inline_equation"
},
{
"bbox": [
434,
553,
505,
566
],
"score": 1.0,
"content": "of all nodes), and",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
396,
577
],
"score": 1.0,
"content": "used the sklearn.neighbors.KNeighborsClassifier with",
"type": "text"
},
{
"bbox": [
396,
565,
430,
575
],
"score": 0.89,
"content": "k = 1 0",
"type": "inline_equation"
},
{
"bbox": [
430,
564,
506,
577
],
"score": 1.0,
"content": "(Pedregosa et al.,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
574,
505,
590
],
"spans": [
{
"bbox": [
105,
574,
505,
590
],
"score": 1.0,
"content": "2011). Of note, we used the train/test split only for training the classifier but not for computing",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
586,
505,
600
],
"spans": [
{
"bbox": [
105,
586,
505,
600
],
"score": 1.0,
"content": "the graph embedding itself. We used sklearn.preprocessing.StandardScaler to stan-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
597,
286,
610
],
"spans": [
{
"bbox": [
106,
597,
286,
610
],
"score": 1.0,
"content": "dardize all features based on the training set.",
"type": "text"
}
],
"index": 24
}
],
"index": 21.5
},
{
"type": "text",
"bbox": [
107,
614,
505,
648
],
"lines": [
{
"bbox": [
105,
614,
505,
627
],
"spans": [
{
"bbox": [
105,
614,
192,
627
],
"score": 1.0,
"content": "For graph CNE with",
"type": "text"
},
{
"bbox": [
193,
615,
230,
625
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
230,
614,
505,
627
],
"score": 1.0,
"content": ", trained using cosine distance, we experimented with using cosine-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
625,
505,
638
],
"spans": [
{
"bbox": [
105,
625,
168,
638
],
"score": 1.0,
"content": "distance-based",
"type": "text"
},
{
"bbox": [
168,
626,
190,
636
],
"score": 0.49,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
190,
625,
505,
638
],
"score": 1.0,
"content": "recall and accuracy, but found that it gave very close results to the Euclidean-",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
636,
408,
649
],
"spans": [
{
"bbox": [
106,
636,
167,
649
],
"score": 1.0,
"content": "distance-based",
"type": "text"
},
{
"bbox": [
168,
637,
189,
647
],
"score": 0.41,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
190,
636,
408,
649
],
"score": 1.0,
"content": "evaluations (all differences below 1 percentage point).",
"type": "text"
}
],
"index": 27
}
],
"index": 26
},
{
"type": "text",
"bbox": [
107,
653,
504,
687
],
"lines": [
{
"bbox": [
106,
654,
505,
665
],
"spans": [
{
"bbox": [
106,
654,
505,
665
],
"score": 1.0,
"content": "For linear accuracy we used the sklearn.linear model.LogisticRegression class",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
664,
505,
676
],
"spans": [
{
"bbox": [
106,
664,
241,
676
],
"score": 1.0,
"content": "with no regularization (penalty",
"type": "text"
},
{
"bbox": [
242,
666,
249,
674
],
"score": 0.38,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
249,
664,
505,
676
],
"score": 1.0,
"content": "None) and otherwise default parameters, and the same train/test",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
676,
353,
688
],
"spans": [
{
"bbox": [
106,
676,
353,
688
],
"score": 1.0,
"content": "split. Features were standardized using StandardScaler.",
"type": "text"
}
],
"index": 30
}
],
"index": 29
},
{
"type": "text",
"bbox": [
108,
699,
504,
732
],
"lines": [
{
"bbox": [
106,
698,
506,
712
],
"spans": [
{
"bbox": [
106,
698,
506,
712
],
"score": 1.0,
"content": "Computing environment All computations were performed on a remote computing server with",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "an Intel Xeon Gold CPU with 16 double-threaded 2.9 Ghz cores, 384 GB of RAM, and an NVIDIA",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
720,
505,
733
],
"spans": [
{
"bbox": [
105,
720,
394,
733
],
"score": 1.0,
"content": "RTX A6000 GPU. GPU training was used for CNE models but not for",
"type": "text"
},
{
"bbox": [
395,
721,
400,
730
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
400,
720,
505,
733
],
"score": 1.0,
"content": "-SNE. Computation times",
"type": "text"
}
],
"index": 33
}
],
"index": 32
}
],
"page_idx": 4,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
302,
750,
309,
763
],
"spans": [
{
"bbox": [
302,
750,
309,
763
],
"score": 1.0,
"content": "5",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
104,
75,
496,
294
],
"blocks": [
{
"type": "image_body",
"bbox": [
104,
75,
496,
294
],
"group_id": 0,
"lines": [
{
"bbox": [
104,
75,
496,
294
],
"spans": [
{
"bbox": [
104,
75,
496,
294
],
"score": 0.973,
"type": "image",
"image_path": "2e5f7f25dc75ee881bccb233b1ff2c656b23adc633b4fd11cb1bd5536ccf62fd.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
104,
75,
496,
148.0
],
"spans": [],
"index": 0
},
{
"bbox": [
104,
148.0,
496,
221.0
],
"spans": [],
"index": 1
},
{
"bbox": [
104,
221.0,
496,
294.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
314,
505,
348
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
314,
504,
326
],
"spans": [
{
"bbox": [
106,
314,
504,
326
],
"score": 1.0,
"content": "Figure 2: Embeddings of the ACO and APH datasets obtained using FDP (Fruchterman & Rein-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
325,
505,
338
],
"spans": [
{
"bbox": [
105,
325,
299,
338
],
"score": 1.0,
"content": "gold, 1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
299,
326,
304,
336
],
"score": 0.72,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
305,
325,
474,
338
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), and our graph",
"type": "text"
},
{
"bbox": [
474,
326,
479,
335
],
"score": 0.58,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
480,
325,
505,
338
],
"score": 1.0,
"content": "-SNE.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
336,
505,
349
],
"spans": [
{
"bbox": [
106,
336,
505,
349
],
"score": 1.0,
"content": "Embeddings in each row were aligned using Procrustes rotation. See Figure A.3 for all six datasets.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
106,
366,
505,
400
],
"lines": [
{
"bbox": [
105,
365,
505,
380
],
"spans": [
{
"bbox": [
105,
365,
505,
380
],
"score": 1.0,
"content": "Performance metrics We evaluated the performance of our methods using three performance",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
378,
505,
391
],
"spans": [
{
"bbox": [
105,
378,
142,
391
],
"score": 1.0,
"content": "metrics:",
"type": "text"
},
{
"bbox": [
142,
379,
149,
388
],
"score": 0.79,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
149,
378,
226,
391
],
"score": 1.0,
"content": "-nearest-neighbors",
"type": "text"
},
{
"bbox": [
226,
378,
252,
389
],
"score": 0.32,
"content": "( k \\mathsf { N N } )",
"type": "inline_equation"
},
{
"bbox": [
253,
378,
281,
391
],
"score": 1.0,
"content": "recall,",
"type": "text"
},
{
"bbox": [
281,
378,
303,
388
],
"score": 0.54,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
303,
378,
505,
391
],
"score": 1.0,
"content": "classification accuracy, and, for high-dimensional",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
106,
388,
279,
402
],
"spans": [
{
"bbox": [
106,
388,
279,
402
],
"score": 1.0,
"content": "embeddings, linear classification accuracy.",
"type": "text"
}
],
"index": 8
}
],
"index": 7,
"bbox_fs": [
105,
365,
505,
402
]
},
{
"type": "text",
"bbox": [
107,
405,
505,
439
],
"lines": [
{
"bbox": [
106,
406,
504,
418
],
"spans": [
{
"bbox": [
106,
406,
124,
418
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
124,
406,
146,
416
],
"score": 0.69,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
406,
504,
418
],
"score": 1.0,
"content": "recall quantifies how well local node neighborhoods are preserved in the embedding. We",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
105,
416,
505,
429
],
"spans": [
{
"bbox": [
105,
416,
505,
429
],
"score": 1.0,
"content": "defined it as the average fraction of each node’s graph neighbors that are among the node’s nearest",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
428,
222,
441
],
"spans": [
{
"bbox": [
106,
428,
222,
441
],
"score": 1.0,
"content": "neighbors in the embedding:",
"type": "text"
}
],
"index": 11
}
],
"index": 10,
"bbox_fs": [
105,
406,
505,
441
]
},
{
"type": "interline_equation",
"bbox": [
227,
443,
382,
478
],
"lines": [
{
"bbox": [
227,
443,
382,
478
],
"spans": [
{
"bbox": [
227,
443,
382,
478
],
"score": 0.95,
"content": "\\mathrm { R e c a l l } = \\frac { 1 } { | \\mathcal { V } | } \\sum _ { i = 1 } ^ { | \\mathcal { V } | } \\frac { \\left| N _ { G } [ i ] \\cap N _ { E , k _ { i } } [ i ] \\right| } { k _ { i } } ,",
"type": "interline_equation",
"image_path": "013acbeb5f8e7077d62b07cab43f72279519c94d2eac7331ea5a660e788ee069.jpg"
}
]
}
],
"index": 12.5,
"virtual_lines": [
{
"bbox": [
227,
443,
382,
460.5
],
"spans": [],
"index": 12
},
{
"bbox": [
227,
460.5,
382,
478.0
],
"spans": [],
"index": 13
}
]
},
{
"type": "text",
"bbox": [
107,
481,
505,
538
],
"lines": [
{
"bbox": [
105,
480,
504,
495
],
"spans": [
{
"bbox": [
105,
480,
133,
495
],
"score": 1.0,
"content": "where",
"type": "text"
},
{
"bbox": [
134,
482,
147,
494
],
"score": 0.9,
"content": "| \\nu |",
"type": "inline_equation"
},
{
"bbox": [
147,
480,
294,
495
],
"score": 1.0,
"content": "is the number of nodes in the graph,",
"type": "text"
},
{
"bbox": [
294,
482,
319,
494
],
"score": 0.92,
"content": "N _ { G } [ i ]",
"type": "inline_equation"
},
{
"bbox": [
319,
480,
390,
495
],
"score": 1.0,
"content": "is the set of node",
"type": "text"
},
{
"bbox": [
390,
483,
395,
492
],
"score": 0.64,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
395,
480,
472,
495
],
"score": 1.0,
"content": "’s graph neighbors,",
"type": "text"
},
{
"bbox": [
472,
482,
504,
493
],
"score": 0.89,
"content": "N _ { E , k } [ i ]",
"type": "inline_equation"
}
],
"index": 14
},
{
"bbox": [
105,
490,
504,
507
],
"spans": [
{
"bbox": [
105,
490,
200,
507
],
"score": 1.0,
"content": "denotes the set of node",
"type": "text"
},
{
"bbox": [
200,
494,
205,
503
],
"score": 0.71,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
205,
490,
212,
507
],
"score": 1.0,
"content": "’s",
"type": "text"
},
{
"bbox": [
213,
493,
220,
503
],
"score": 0.8,
"content": "k",
"type": "inline_equation"
},
{
"bbox": [
220,
490,
451,
507
],
"score": 1.0,
"content": "Euclidean nearest neighbors in the embedding space, and",
"type": "text"
},
{
"bbox": [
452,
493,
504,
505
],
"score": 0.9,
"content": "k _ { i } = | N _ { G } [ i ] |",
"type": "inline_equation"
}
],
"index": 15
},
{
"bbox": [
105,
504,
505,
516
],
"spans": [
{
"bbox": [
105,
504,
197,
516
],
"score": 1.0,
"content": "is the number of node",
"type": "text"
},
{
"bbox": [
198,
505,
203,
514
],
"score": 0.71,
"content": "i",
"type": "inline_equation"
},
{
"bbox": [
203,
504,
505,
516
],
"score": 1.0,
"content": "’s graph neighbors. This metric is similar to what is commonly used in the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
514,
506,
528
],
"spans": [
{
"bbox": [
105,
514,
506,
528
],
"score": 1.0,
"content": "literature to benchmark graph layout algorithms (Kruiger et al., 2017; Zhu et al., 2020a; Zhong et al.,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
106,
525,
380,
539
],
"spans": [
{
"bbox": [
106,
525,
380,
539
],
"score": 1.0,
"content": "2023), and so is our main metric for measuring graph layout quality.",
"type": "text"
}
],
"index": 18
}
],
"index": 16,
"bbox_fs": [
105,
480,
506,
539
]
},
{
"type": "text",
"bbox": [
107,
542,
505,
609
],
"lines": [
{
"bbox": [
106,
542,
505,
555
],
"spans": [
{
"bbox": [
106,
542,
125,
555
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
125,
543,
147,
553
],
"score": 0.31,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
542,
505,
555
],
"score": 1.0,
"content": "classification accuracy quantifies local class separation in the embedding. To calculate",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
107,
553,
505,
566
],
"spans": [
{
"bbox": [
107,
554,
128,
564
],
"score": 0.51,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
128,
553,
297,
566
],
"score": 1.0,
"content": "accuracy, we split all nodes into a training",
"type": "text"
},
{
"bbox": [
298,
554,
312,
566
],
"score": 0.8,
"content": "( 2 / 3",
"type": "inline_equation"
},
{
"bbox": [
313,
553,
417,
566
],
"score": 1.0,
"content": "of all nodes) and a test set",
"type": "text"
},
{
"bbox": [
418,
554,
433,
565
],
"score": 0.83,
"content": "( 1 / 3",
"type": "inline_equation"
},
{
"bbox": [
434,
553,
505,
566
],
"score": 1.0,
"content": "of all nodes), and",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
564,
506,
577
],
"spans": [
{
"bbox": [
105,
564,
396,
577
],
"score": 1.0,
"content": "used the sklearn.neighbors.KNeighborsClassifier with",
"type": "text"
},
{
"bbox": [
396,
565,
430,
575
],
"score": 0.89,
"content": "k = 1 0",
"type": "inline_equation"
},
{
"bbox": [
430,
564,
506,
577
],
"score": 1.0,
"content": "(Pedregosa et al.,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
574,
505,
590
],
"spans": [
{
"bbox": [
105,
574,
505,
590
],
"score": 1.0,
"content": "2011). Of note, we used the train/test split only for training the classifier but not for computing",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
586,
505,
600
],
"spans": [
{
"bbox": [
105,
586,
505,
600
],
"score": 1.0,
"content": "the graph embedding itself. We used sklearn.preprocessing.StandardScaler to stan-",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
106,
597,
286,
610
],
"spans": [
{
"bbox": [
106,
597,
286,
610
],
"score": 1.0,
"content": "dardize all features based on the training set.",
"type": "text"
}
],
"index": 24
}
],
"index": 21.5,
"bbox_fs": [
105,
542,
506,
610
]
},
{
"type": "text",
"bbox": [
107,
614,
505,
648
],
"lines": [
{
"bbox": [
105,
614,
505,
627
],
"spans": [
{
"bbox": [
105,
614,
192,
627
],
"score": 1.0,
"content": "For graph CNE with",
"type": "text"
},
{
"bbox": [
193,
615,
230,
625
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
230,
614,
505,
627
],
"score": 1.0,
"content": ", trained using cosine distance, we experimented with using cosine-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
625,
505,
638
],
"spans": [
{
"bbox": [
105,
625,
168,
638
],
"score": 1.0,
"content": "distance-based",
"type": "text"
},
{
"bbox": [
168,
626,
190,
636
],
"score": 0.49,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
190,
625,
505,
638
],
"score": 1.0,
"content": "recall and accuracy, but found that it gave very close results to the Euclidean-",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
636,
408,
649
],
"spans": [
{
"bbox": [
106,
636,
167,
649
],
"score": 1.0,
"content": "distance-based",
"type": "text"
},
{
"bbox": [
168,
637,
189,
647
],
"score": 0.41,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
190,
636,
408,
649
],
"score": 1.0,
"content": "evaluations (all differences below 1 percentage point).",
"type": "text"
}
],
"index": 27
}
],
"index": 26,
"bbox_fs": [
105,
614,
505,
649
]
},
{
"type": "text",
"bbox": [
107,
653,
504,
687
],
"lines": [
{
"bbox": [
106,
654,
505,
665
],
"spans": [
{
"bbox": [
106,
654,
505,
665
],
"score": 1.0,
"content": "For linear accuracy we used the sklearn.linear model.LogisticRegression class",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
664,
505,
676
],
"spans": [
{
"bbox": [
106,
664,
241,
676
],
"score": 1.0,
"content": "with no regularization (penalty",
"type": "text"
},
{
"bbox": [
242,
666,
249,
674
],
"score": 0.38,
"content": "=",
"type": "inline_equation"
},
{
"bbox": [
249,
664,
505,
676
],
"score": 1.0,
"content": "None) and otherwise default parameters, and the same train/test",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
676,
353,
688
],
"spans": [
{
"bbox": [
106,
676,
353,
688
],
"score": 1.0,
"content": "split. Features were standardized using StandardScaler.",
"type": "text"
}
],
"index": 30
}
],
"index": 29,
"bbox_fs": [
106,
654,
505,
688
]
},
{
"type": "text",
"bbox": [
108,
699,
504,
732
],
"lines": [
{
"bbox": [
106,
698,
506,
712
],
"spans": [
{
"bbox": [
106,
698,
506,
712
],
"score": 1.0,
"content": "Computing environment All computations were performed on a remote computing server with",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
709,
505,
722
],
"spans": [
{
"bbox": [
106,
709,
505,
722
],
"score": 1.0,
"content": "an Intel Xeon Gold CPU with 16 double-threaded 2.9 Ghz cores, 384 GB of RAM, and an NVIDIA",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
720,
505,
733
],
"spans": [
{
"bbox": [
105,
720,
394,
733
],
"score": 1.0,
"content": "RTX A6000 GPU. GPU training was used for CNE models but not for",
"type": "text"
},
{
"bbox": [
395,
721,
400,
730
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
400,
720,
505,
733
],
"score": 1.0,
"content": "-SNE. Computation times",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
278,
505,
290
],
"spans": [
{
"bbox": [
106,
278,
357,
290
],
"score": 1.0,
"content": "are shown in Figure A.1. For the largest dataset (ARX), graph",
"type": "text",
"cross_page": true
},
{
"bbox": [
357,
279,
362,
288
],
"score": 0.66,
"content": "t",
"type": "inline_equation",
"cross_page": true
},
{
"bbox": [
363,
278,
505,
290
],
"score": 1.0,
"content": "-SNE took around 100 seconds and",
"type": "text",
"cross_page": true
}
],
"index": 5
},
{
"bbox": [
105,
289,
253,
302
],
"spans": [
{
"bbox": [
105,
289,
253,
302
],
"score": 1.0,
"content": "graph CNE took around 60 minutes.",
"type": "text",
"cross_page": true
}
],
"index": 6
}
],
"index": 32,
"bbox_fs": [
105,
698,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
107,
81,
504,
225
],
"blocks": [
{
"type": "image_body",
"bbox": [
107,
81,
504,
225
],
"group_id": 0,
"lines": [
{
"bbox": [
107,
81,
504,
225
],
"spans": [
{
"bbox": [
107,
81,
504,
225
],
"score": 0.968,
"type": "image",
"image_path": "36d02dd4d58e27a78df1813d2a0ebc6d57cdb0d6a908f0caf5fd9adffc64418a.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
107,
81,
504,
129.0
],
"spans": [],
"index": 0
},
{
"bbox": [
107,
129.0,
504,
177.0
],
"spans": [],
"index": 1
},
{
"bbox": [
107,
177.0,
504,
225.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
237,
504,
260
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
237,
505,
250
],
"spans": [
{
"bbox": [
105,
237,
316,
250
],
"score": 1.0,
"content": "Figure 3: Performance metrics for graph layouts:",
"type": "text"
},
{
"bbox": [
317,
237,
339,
248
],
"score": 0.34,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
339,
237,
384,
250
],
"score": 1.0,
"content": "recall and",
"type": "text"
},
{
"bbox": [
385,
238,
407,
248
],
"score": 0.67,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
407,
237,
505,
250
],
"score": 1.0,
"content": "accuracy. Datasets are",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
249,
474,
261
],
"spans": [
{
"bbox": [
106,
249,
474,
261
],
"score": 1.0,
"content": "ordered by the increasing sample size. See Figures 2 and A.3 for the corresponding layouts.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
106,
278,
504,
300
],
"lines": [
{
"bbox": [
106,
278,
505,
290
],
"spans": [
{
"bbox": [
106,
278,
357,
290
],
"score": 1.0,
"content": "are shown in Figure A.1. For the largest dataset (ARX), graph",
"type": "text"
},
{
"bbox": [
357,
279,
362,
288
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
363,
278,
505,
290
],
"score": 1.0,
"content": "-SNE took around 100 seconds and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
289,
253,
302
],
"spans": [
{
"bbox": [
105,
289,
253,
302
],
"score": 1.0,
"content": "graph CNE took around 60 minutes.",
"type": "text"
}
],
"index": 6
}
],
"index": 5.5
},
{
"type": "title",
"bbox": [
107,
316,
311,
329
],
"lines": [
{
"bbox": [
105,
316,
311,
330
],
"spans": [
{
"bbox": [
105,
316,
275,
330
],
"score": 1.0,
"content": "5 GRAPH LAYOUTS VIA GRAPH",
"type": "text"
},
{
"bbox": [
276,
317,
281,
327
],
"score": 0.59,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
282,
316,
311,
330
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
106,
340,
505,
385
],
"lines": [
{
"bbox": [
105,
340,
505,
354
],
"spans": [
{
"bbox": [
105,
340,
124,
354
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
125,
342,
130,
351
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
131,
340,
505,
354
],
"score": 1.0,
"content": "-SNE algorithm consists of two steps: first, it computes pairwise affinities between all pairs",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
352,
506,
365
],
"spans": [
{
"bbox": [
105,
352,
202,
365
],
"score": 1.0,
"content": "of points based on the",
"type": "text"
},
{
"bbox": [
202,
352,
224,
362
],
"score": 0.73,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
224,
352,
506,
365
],
"score": 1.0,
"content": "graph; second, it optimizes the embedding to match these affinities",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
363,
505,
375
],
"spans": [
{
"bbox": [
106,
363,
204,
375
],
"score": 1.0,
"content": "(Section 3.1). For graph",
"type": "text"
},
{
"bbox": [
205,
364,
210,
373
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
210,
363,
505,
375
],
"score": 1.0,
"content": "-SNE, we replace the first step and obtain the affinity matrix directly from",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
374,
503,
386
],
"spans": [
{
"bbox": [
106,
374,
271,
386
],
"score": 1.0,
"content": "the graph adjacency matrix. We then run",
"type": "text"
},
{
"bbox": [
271,
375,
276,
384
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
276,
374,
503,
386
],
"score": 1.0,
"content": "-SNE optimization to produce the embedding (Figure 1).",
"type": "text"
}
],
"index": 11
}
],
"index": 9.5
},
{
"type": "text",
"bbox": [
106,
390,
505,
446
],
"lines": [
{
"bbox": [
106,
390,
506,
403
],
"spans": [
{
"bbox": [
106,
390,
222,
403
],
"score": 1.0,
"content": "Given an unweighted graph",
"type": "text"
},
{
"bbox": [
222,
390,
273,
403
],
"score": 0.94,
"content": "G = ( \\nu , \\mathcal { E } )",
"type": "inline_equation"
},
{
"bbox": [
273,
390,
362,
403
],
"score": 1.0,
"content": ", its adjacency matrix",
"type": "text"
},
{
"bbox": [
363,
391,
373,
401
],
"score": 0.31,
"content": "\\mathbf { A }",
"type": "inline_equation"
},
{
"bbox": [
373,
390,
457,
403
],
"score": 1.0,
"content": "is defined such that",
"type": "text"
},
{
"bbox": [
457,
391,
495,
403
],
"score": 0.93,
"content": "A _ { i j } = 1",
"type": "inline_equation"
},
{
"bbox": [
495,
390,
506,
403
],
"score": 1.0,
"content": "if",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
107,
402,
506,
415
],
"spans": [
{
"bbox": [
107,
402,
150,
414
],
"score": 0.92,
"content": "( i , j ) \\in \\mathcal { E }",
"type": "inline_equation"
},
{
"bbox": [
150,
402,
169,
415
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
170,
402,
208,
414
],
"score": 0.93,
"content": "A _ { i j } = 0",
"type": "inline_equation"
},
{
"bbox": [
208,
402,
506,
415
],
"score": 1.0,
"content": "otherwise. Since all graphs considered in this study are undirected, the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
412,
505,
426
],
"spans": [
{
"bbox": [
105,
412,
298,
426
],
"score": 1.0,
"content": "adjacency matrix is a binary, symmetric square",
"type": "text"
},
{
"bbox": [
298,
414,
324,
423
],
"score": 0.9,
"content": "n \\times n",
"type": "inline_equation"
},
{
"bbox": [
324,
412,
505,
426
],
"score": 1.0,
"content": "matrix. In order to convert it into an affinity",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
422,
505,
436
],
"spans": [
{
"bbox": [
105,
422,
182,
436
],
"score": 1.0,
"content": "matrix suitable for",
"type": "text"
},
{
"bbox": [
183,
424,
188,
433
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
188,
422,
307,
436
],
"score": 1.0,
"content": "-SNE, we follow the standard",
"type": "text"
},
{
"bbox": [
307,
424,
312,
433
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
313,
422,
505,
436
],
"score": 1.0,
"content": "-SNE’s approach (Section 3.1): divide each row",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
434,
497,
448
],
"spans": [
{
"bbox": [
106,
434,
497,
448
],
"score": 1.0,
"content": "by the sum of its elements, then symmetrize the resulting matrix, and then normalize to sum to 1:",
"type": "text"
}
],
"index": 16
}
],
"index": 14
},
{
"type": "interline_equation",
"bbox": [
213,
448,
396,
482
],
"lines": [
{
"bbox": [
213,
448,
396,
482
],
"spans": [
{
"bbox": [
213,
448,
396,
482
],
"score": 0.9,
"content": "\\mathbf { P } = { \\frac { { \\tilde { \\mathbf { A } } } + { \\tilde { \\mathbf { A } } } ^ { \\top } } { 2 n } } , { \\mathrm { ~ w h e r e ~ } } { \\tilde { A } } _ { i j } = A _ { i j } { \\Big / } \\sum _ { k = 1 } ^ { n } A _ { i k } .",
"type": "interline_equation",
"image_path": "059da263f07d66cdd192e32bb1681c7b7d46a29444d542c16284af5f58184c25.jpg"
}
]
}
],
"index": 17.5,
"virtual_lines": [
{
"bbox": [
213,
448,
396,
465.0
],
"spans": [],
"index": 17
},
{
"bbox": [
213,
465.0,
396,
482.0
],
"spans": [],
"index": 18
}
]
},
{
"type": "text",
"bbox": [
107,
488,
505,
545
],
"lines": [
{
"bbox": [
105,
488,
506,
502
],
"spans": [
{
"bbox": [
105,
488,
506,
502
],
"score": 1.0,
"content": "For optimization, we used the openTSNE library (Policar et al., 2019) with default parameters. It ˇ",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "uses Laplacian Eigenmaps (Belkin & Niyogi, 2003) for initialization (Kobak & Linderman, 2021),",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
510,
505,
524
],
"spans": [
{
"bbox": [
105,
510,
230,
524
],
"score": 1.0,
"content": "sets the learning rate equal to",
"type": "text"
},
{
"bbox": [
230,
514,
237,
521
],
"score": 0.74,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
238,
510,
505,
524
],
"score": 1.0,
"content": "to achieve good convergence (Linderman & Steinerberger, 2019;",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
521,
505,
536
],
"spans": [
{
"bbox": [
105,
521,
398,
536
],
"score": 1.0,
"content": "Belkina et al., 2019), and employs fast FIt-SNE algorithm that has linear",
"type": "text"
},
{
"bbox": [
398,
522,
421,
534
],
"score": 0.92,
"content": "{ \\mathcal { O } } ( n )",
"type": "inline_equation"
},
{
"bbox": [
421,
521,
505,
536
],
"score": 1.0,
"content": "runtime (Linderman",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
533,
159,
546
],
"spans": [
{
"bbox": [
105,
533,
159,
546
],
"score": 1.0,
"content": "et al., 2019).",
"type": "text"
}
],
"index": 23
}
],
"index": 21
},
{
"type": "text",
"bbox": [
106,
549,
505,
639
],
"lines": [
{
"bbox": [
106,
549,
505,
563
],
"spans": [
{
"bbox": [
106,
549,
188,
563
],
"score": 1.0,
"content": "We compared graph",
"type": "text"
},
{
"bbox": [
189,
551,
193,
560
],
"score": 0.59,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
194,
549,
505,
563
],
"score": 1.0,
"content": "-SNE with three existing graph layout algorithms: FDP (Fruchterman & Rein-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
561,
506,
574
],
"spans": [
{
"bbox": [
105,
561,
304,
574
],
"score": 1.0,
"content": "gold, 1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
304,
562,
309,
571
],
"score": 0.75,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
309,
561,
506,
574
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023). We chose FDP be-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "cause it is the default layout algorithm in a popular NetworkX package (Hagberg et al., 2008).",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
583,
505,
595
],
"spans": [
{
"bbox": [
106,
583,
199,
595
],
"score": 1.0,
"content": "Two other algorithms,",
"type": "text"
},
{
"bbox": [
199,
584,
204,
593
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
204,
583,
505,
595
],
"score": 1.0,
"content": "-FDP and DRGraph, are recent and can be considered state-of-the-art (we",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
594,
504,
606
],
"spans": [
{
"bbox": [
106,
594,
504,
606
],
"score": 1.0,
"content": "did not use tsNET (Kruiger et al., 2017) for benchmarking, because it cannot embed large graphs",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
605,
506,
617
],
"spans": [
{
"bbox": [
105,
605,
506,
617
],
"score": 1.0,
"content": "and is outperformed by its successor DRGraph). We used the NetworkX implementation of FDP",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
616,
505,
628
],
"spans": [
{
"bbox": [
106,
616,
505,
628
],
"score": 1.0,
"content": "(networkx.drawing.layout.spring layout) and the original implementations of both",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
626,
349,
640
],
"spans": [
{
"bbox": [
106,
628,
111,
637
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
112,
626,
349,
640
],
"score": 1.0,
"content": "-FDP and DRGraph, all with default parameters (Figure 2).",
"type": "text"
}
],
"index": 31
}
],
"index": 27.5
},
{
"type": "text",
"bbox": [
106,
643,
505,
732
],
"lines": [
{
"bbox": [
106,
643,
505,
657
],
"spans": [
{
"bbox": [
106,
643,
191,
657
],
"score": 1.0,
"content": "We found that graph",
"type": "text"
},
{
"bbox": [
192,
645,
196,
654
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
197,
643,
457,
657
],
"score": 1.0,
"content": "-SNE consistently outperformed all competitors in terms of both",
"type": "text"
},
{
"bbox": [
458,
644,
479,
654
],
"score": 0.36,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
480,
643,
505,
657
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
655,
505,
666
],
"spans": [
{
"bbox": [
106,
655,
124,
666
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
124,
655,
146,
666
],
"score": 0.71,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
655,
505,
666
],
"score": 1.0,
"content": "accuracy (Figure 3): it showed the highest values on all datasets, 12 out of 12 times.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
666,
506,
679
],
"spans": [
{
"bbox": [
105,
666,
506,
679
],
"score": 1.0,
"content": "In agreement with the original results of Zhu et al. (2020a) and Zhong et al. (2023), we saw that",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
164,
689
],
"score": 1.0,
"content": "DRGraph and",
"type": "text"
},
{
"bbox": [
164,
677,
169,
687
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
169,
677,
376,
689
],
"score": 1.0,
"content": "-FDP outperformed FDP in both metrics. Our graph",
"type": "text"
},
{
"bbox": [
376,
678,
381,
687
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
381,
677,
505,
689
],
"score": 1.0,
"content": "-SNE showed further improve-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
687,
505,
701
],
"spans": [
{
"bbox": [
105,
687,
269,
701
],
"score": 1.0,
"content": "ment, and it was substantial: in terms of",
"type": "text"
},
{
"bbox": [
270,
688,
291,
698
],
"score": 0.57,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
291,
687,
344,
701
],
"score": 1.0,
"content": "recall, graph",
"type": "text"
},
{
"bbox": [
344,
689,
349,
698
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
350,
687,
505,
701
],
"score": 1.0,
"content": "-SNE improved on the best competitor",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
698,
505,
713
],
"spans": [
{
"bbox": [
105,
698,
317,
713
],
"score": 1.0,
"content": "on average by 18.2 percentage points, and in terms of",
"type": "text"
},
{
"bbox": [
317,
699,
339,
709
],
"score": 0.35,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
339,
698,
505,
713
],
"score": 1.0,
"content": "accuracy — on average by 6.7 percentage",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
710,
505,
722
],
"spans": [
{
"bbox": [
105,
710,
505,
722
],
"score": 1.0,
"content": "points. The improvement was particularly strong for the largest graph (ARX), where performance",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
721,
264,
733
],
"spans": [
{
"bbox": [
106,
721,
264,
733
],
"score": 1.0,
"content": "of other methods strongly deteriorated.",
"type": "text"
}
],
"index": 39
}
],
"index": 35.5
}
],
"page_idx": 5,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
752,
309,
760
],
"lines": [
{
"bbox": [
302,
751,
310,
762
],
"spans": [
{
"bbox": [
302,
751,
310,
762
],
"score": 1.0,
"content": "6",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
107,
81,
504,
225
],
"blocks": [
{
"type": "image_body",
"bbox": [
107,
81,
504,
225
],
"group_id": 0,
"lines": [
{
"bbox": [
107,
81,
504,
225
],
"spans": [
{
"bbox": [
107,
81,
504,
225
],
"score": 0.968,
"type": "image",
"image_path": "36d02dd4d58e27a78df1813d2a0ebc6d57cdb0d6a908f0caf5fd9adffc64418a.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
107,
81,
504,
129.0
],
"spans": [],
"index": 0
},
{
"bbox": [
107,
129.0,
504,
177.0
],
"spans": [],
"index": 1
},
{
"bbox": [
107,
177.0,
504,
225.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
237,
504,
260
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
237,
505,
250
],
"spans": [
{
"bbox": [
105,
237,
316,
250
],
"score": 1.0,
"content": "Figure 3: Performance metrics for graph layouts:",
"type": "text"
},
{
"bbox": [
317,
237,
339,
248
],
"score": 0.34,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
339,
237,
384,
250
],
"score": 1.0,
"content": "recall and",
"type": "text"
},
{
"bbox": [
385,
238,
407,
248
],
"score": 0.67,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
407,
237,
505,
250
],
"score": 1.0,
"content": "accuracy. Datasets are",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
249,
474,
261
],
"spans": [
{
"bbox": [
106,
249,
474,
261
],
"score": 1.0,
"content": "ordered by the increasing sample size. See Figures 2 and A.3 for the corresponding layouts.",
"type": "text"
}
],
"index": 4
}
],
"index": 3.5
}
],
"index": 2.25
},
{
"type": "text",
"bbox": [
106,
278,
504,
300
],
"lines": [],
"index": 5.5,
"bbox_fs": [
105,
278,
505,
302
],
"lines_deleted": true
},
{
"type": "title",
"bbox": [
107,
316,
311,
329
],
"lines": [
{
"bbox": [
105,
316,
311,
330
],
"spans": [
{
"bbox": [
105,
316,
275,
330
],
"score": 1.0,
"content": "5 GRAPH LAYOUTS VIA GRAPH",
"type": "text"
},
{
"bbox": [
276,
317,
281,
327
],
"score": 0.59,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
282,
316,
311,
330
],
"score": 1.0,
"content": "-SNE",
"type": "text"
}
],
"index": 7
}
],
"index": 7
},
{
"type": "text",
"bbox": [
106,
340,
505,
385
],
"lines": [
{
"bbox": [
105,
340,
505,
354
],
"spans": [
{
"bbox": [
105,
340,
124,
354
],
"score": 1.0,
"content": "The",
"type": "text"
},
{
"bbox": [
125,
342,
130,
351
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
131,
340,
505,
354
],
"score": 1.0,
"content": "-SNE algorithm consists of two steps: first, it computes pairwise affinities between all pairs",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
352,
506,
365
],
"spans": [
{
"bbox": [
105,
352,
202,
365
],
"score": 1.0,
"content": "of points based on the",
"type": "text"
},
{
"bbox": [
202,
352,
224,
362
],
"score": 0.73,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
224,
352,
506,
365
],
"score": 1.0,
"content": "graph; second, it optimizes the embedding to match these affinities",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
363,
505,
375
],
"spans": [
{
"bbox": [
106,
363,
204,
375
],
"score": 1.0,
"content": "(Section 3.1). For graph",
"type": "text"
},
{
"bbox": [
205,
364,
210,
373
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
210,
363,
505,
375
],
"score": 1.0,
"content": "-SNE, we replace the first step and obtain the affinity matrix directly from",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
106,
374,
503,
386
],
"spans": [
{
"bbox": [
106,
374,
271,
386
],
"score": 1.0,
"content": "the graph adjacency matrix. We then run",
"type": "text"
},
{
"bbox": [
271,
375,
276,
384
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
276,
374,
503,
386
],
"score": 1.0,
"content": "-SNE optimization to produce the embedding (Figure 1).",
"type": "text"
}
],
"index": 11
}
],
"index": 9.5,
"bbox_fs": [
105,
340,
506,
386
]
},
{
"type": "text",
"bbox": [
106,
390,
505,
446
],
"lines": [
{
"bbox": [
106,
390,
506,
403
],
"spans": [
{
"bbox": [
106,
390,
222,
403
],
"score": 1.0,
"content": "Given an unweighted graph",
"type": "text"
},
{
"bbox": [
222,
390,
273,
403
],
"score": 0.94,
"content": "G = ( \\nu , \\mathcal { E } )",
"type": "inline_equation"
},
{
"bbox": [
273,
390,
362,
403
],
"score": 1.0,
"content": ", its adjacency matrix",
"type": "text"
},
{
"bbox": [
363,
391,
373,
401
],
"score": 0.31,
"content": "\\mathbf { A }",
"type": "inline_equation"
},
{
"bbox": [
373,
390,
457,
403
],
"score": 1.0,
"content": "is defined such that",
"type": "text"
},
{
"bbox": [
457,
391,
495,
403
],
"score": 0.93,
"content": "A _ { i j } = 1",
"type": "inline_equation"
},
{
"bbox": [
495,
390,
506,
403
],
"score": 1.0,
"content": "if",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
107,
402,
506,
415
],
"spans": [
{
"bbox": [
107,
402,
150,
414
],
"score": 0.92,
"content": "( i , j ) \\in \\mathcal { E }",
"type": "inline_equation"
},
{
"bbox": [
150,
402,
169,
415
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
170,
402,
208,
414
],
"score": 0.93,
"content": "A _ { i j } = 0",
"type": "inline_equation"
},
{
"bbox": [
208,
402,
506,
415
],
"score": 1.0,
"content": "otherwise. Since all graphs considered in this study are undirected, the",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
412,
505,
426
],
"spans": [
{
"bbox": [
105,
412,
298,
426
],
"score": 1.0,
"content": "adjacency matrix is a binary, symmetric square",
"type": "text"
},
{
"bbox": [
298,
414,
324,
423
],
"score": 0.9,
"content": "n \\times n",
"type": "inline_equation"
},
{
"bbox": [
324,
412,
505,
426
],
"score": 1.0,
"content": "matrix. In order to convert it into an affinity",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
422,
505,
436
],
"spans": [
{
"bbox": [
105,
422,
182,
436
],
"score": 1.0,
"content": "matrix suitable for",
"type": "text"
},
{
"bbox": [
183,
424,
188,
433
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
188,
422,
307,
436
],
"score": 1.0,
"content": "-SNE, we follow the standard",
"type": "text"
},
{
"bbox": [
307,
424,
312,
433
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
313,
422,
505,
436
],
"score": 1.0,
"content": "-SNE’s approach (Section 3.1): divide each row",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
106,
434,
497,
448
],
"spans": [
{
"bbox": [
106,
434,
497,
448
],
"score": 1.0,
"content": "by the sum of its elements, then symmetrize the resulting matrix, and then normalize to sum to 1:",
"type": "text"
}
],
"index": 16
}
],
"index": 14,
"bbox_fs": [
105,
390,
506,
448
]
},
{
"type": "interline_equation",
"bbox": [
213,
448,
396,
482
],
"lines": [
{
"bbox": [
213,
448,
396,
482
],
"spans": [
{
"bbox": [
213,
448,
396,
482
],
"score": 0.9,
"content": "\\mathbf { P } = { \\frac { { \\tilde { \\mathbf { A } } } + { \\tilde { \\mathbf { A } } } ^ { \\top } } { 2 n } } , { \\mathrm { ~ w h e r e ~ } } { \\tilde { A } } _ { i j } = A _ { i j } { \\Big / } \\sum _ { k = 1 } ^ { n } A _ { i k } .",
"type": "interline_equation",
"image_path": "059da263f07d66cdd192e32bb1681c7b7d46a29444d542c16284af5f58184c25.jpg"
}
]
}
],
"index": 17.5,
"virtual_lines": [
{
"bbox": [
213,
448,
396,
465.0
],
"spans": [],
"index": 17
},
{
"bbox": [
213,
465.0,
396,
482.0
],
"spans": [],
"index": 18
}
]
},
{
"type": "text",
"bbox": [
107,
488,
505,
545
],
"lines": [
{
"bbox": [
105,
488,
506,
502
],
"spans": [
{
"bbox": [
105,
488,
506,
502
],
"score": 1.0,
"content": "For optimization, we used the openTSNE library (Policar et al., 2019) with default parameters. It ˇ",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
501,
505,
513
],
"spans": [
{
"bbox": [
106,
501,
505,
513
],
"score": 1.0,
"content": "uses Laplacian Eigenmaps (Belkin & Niyogi, 2003) for initialization (Kobak & Linderman, 2021),",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
510,
505,
524
],
"spans": [
{
"bbox": [
105,
510,
230,
524
],
"score": 1.0,
"content": "sets the learning rate equal to",
"type": "text"
},
{
"bbox": [
230,
514,
237,
521
],
"score": 0.74,
"content": "n",
"type": "inline_equation"
},
{
"bbox": [
238,
510,
505,
524
],
"score": 1.0,
"content": "to achieve good convergence (Linderman & Steinerberger, 2019;",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
521,
505,
536
],
"spans": [
{
"bbox": [
105,
521,
398,
536
],
"score": 1.0,
"content": "Belkina et al., 2019), and employs fast FIt-SNE algorithm that has linear",
"type": "text"
},
{
"bbox": [
398,
522,
421,
534
],
"score": 0.92,
"content": "{ \\mathcal { O } } ( n )",
"type": "inline_equation"
},
{
"bbox": [
421,
521,
505,
536
],
"score": 1.0,
"content": "runtime (Linderman",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
533,
159,
546
],
"spans": [
{
"bbox": [
105,
533,
159,
546
],
"score": 1.0,
"content": "et al., 2019).",
"type": "text"
}
],
"index": 23
}
],
"index": 21,
"bbox_fs": [
105,
488,
506,
546
]
},
{
"type": "text",
"bbox": [
106,
549,
505,
639
],
"lines": [
{
"bbox": [
106,
549,
505,
563
],
"spans": [
{
"bbox": [
106,
549,
188,
563
],
"score": 1.0,
"content": "We compared graph",
"type": "text"
},
{
"bbox": [
189,
551,
193,
560
],
"score": 0.59,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
194,
549,
505,
563
],
"score": 1.0,
"content": "-SNE with three existing graph layout algorithms: FDP (Fruchterman & Rein-",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
561,
506,
574
],
"spans": [
{
"bbox": [
105,
561,
304,
574
],
"score": 1.0,
"content": "gold, 1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
304,
562,
309,
571
],
"score": 0.75,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
309,
561,
506,
574
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023). We chose FDP be-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
572,
505,
585
],
"spans": [
{
"bbox": [
105,
572,
505,
585
],
"score": 1.0,
"content": "cause it is the default layout algorithm in a popular NetworkX package (Hagberg et al., 2008).",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
583,
505,
595
],
"spans": [
{
"bbox": [
106,
583,
199,
595
],
"score": 1.0,
"content": "Two other algorithms,",
"type": "text"
},
{
"bbox": [
199,
584,
204,
593
],
"score": 0.71,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
204,
583,
505,
595
],
"score": 1.0,
"content": "-FDP and DRGraph, are recent and can be considered state-of-the-art (we",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
106,
594,
504,
606
],
"spans": [
{
"bbox": [
106,
594,
504,
606
],
"score": 1.0,
"content": "did not use tsNET (Kruiger et al., 2017) for benchmarking, because it cannot embed large graphs",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
605,
506,
617
],
"spans": [
{
"bbox": [
105,
605,
506,
617
],
"score": 1.0,
"content": "and is outperformed by its successor DRGraph). We used the NetworkX implementation of FDP",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
616,
505,
628
],
"spans": [
{
"bbox": [
106,
616,
505,
628
],
"score": 1.0,
"content": "(networkx.drawing.layout.spring layout) and the original implementations of both",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
106,
626,
349,
640
],
"spans": [
{
"bbox": [
106,
628,
111,
637
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
112,
626,
349,
640
],
"score": 1.0,
"content": "-FDP and DRGraph, all with default parameters (Figure 2).",
"type": "text"
}
],
"index": 31
}
],
"index": 27.5,
"bbox_fs": [
105,
549,
506,
640
]
},
{
"type": "text",
"bbox": [
106,
643,
505,
732
],
"lines": [
{
"bbox": [
106,
643,
505,
657
],
"spans": [
{
"bbox": [
106,
643,
191,
657
],
"score": 1.0,
"content": "We found that graph",
"type": "text"
},
{
"bbox": [
192,
645,
196,
654
],
"score": 0.63,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
197,
643,
457,
657
],
"score": 1.0,
"content": "-SNE consistently outperformed all competitors in terms of both",
"type": "text"
},
{
"bbox": [
458,
644,
479,
654
],
"score": 0.36,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
480,
643,
505,
657
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
655,
505,
666
],
"spans": [
{
"bbox": [
106,
655,
124,
666
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
124,
655,
146,
666
],
"score": 0.71,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
147,
655,
505,
666
],
"score": 1.0,
"content": "accuracy (Figure 3): it showed the highest values on all datasets, 12 out of 12 times.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
666,
506,
679
],
"spans": [
{
"bbox": [
105,
666,
506,
679
],
"score": 1.0,
"content": "In agreement with the original results of Zhu et al. (2020a) and Zhong et al. (2023), we saw that",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
106,
677,
505,
689
],
"spans": [
{
"bbox": [
106,
677,
164,
689
],
"score": 1.0,
"content": "DRGraph and",
"type": "text"
},
{
"bbox": [
164,
677,
169,
687
],
"score": 0.77,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
169,
677,
376,
689
],
"score": 1.0,
"content": "-FDP outperformed FDP in both metrics. Our graph",
"type": "text"
},
{
"bbox": [
376,
678,
381,
687
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
381,
677,
505,
689
],
"score": 1.0,
"content": "-SNE showed further improve-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
687,
505,
701
],
"spans": [
{
"bbox": [
105,
687,
269,
701
],
"score": 1.0,
"content": "ment, and it was substantial: in terms of",
"type": "text"
},
{
"bbox": [
270,
688,
291,
698
],
"score": 0.57,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
291,
687,
344,
701
],
"score": 1.0,
"content": "recall, graph",
"type": "text"
},
{
"bbox": [
344,
689,
349,
698
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
350,
687,
505,
701
],
"score": 1.0,
"content": "-SNE improved on the best competitor",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
698,
505,
713
],
"spans": [
{
"bbox": [
105,
698,
317,
713
],
"score": 1.0,
"content": "on average by 18.2 percentage points, and in terms of",
"type": "text"
},
{
"bbox": [
317,
699,
339,
709
],
"score": 0.35,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
339,
698,
505,
713
],
"score": 1.0,
"content": "accuracy — on average by 6.7 percentage",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
710,
505,
722
],
"spans": [
{
"bbox": [
105,
710,
505,
722
],
"score": 1.0,
"content": "points. The improvement was particularly strong for the largest graph (ARX), where performance",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
721,
264,
733
],
"spans": [
{
"bbox": [
106,
721,
264,
733
],
"score": 1.0,
"content": "of other methods strongly deteriorated.",
"type": "text"
}
],
"index": 39
}
],
"index": 35.5,
"bbox_fs": [
105,
643,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
108,
81,
504,
225
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
81,
504,
225
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
81,
504,
225
],
"spans": [
{
"bbox": [
108,
81,
504,
225
],
"score": 0.968,
"type": "image",
"image_path": "472c51903edfd06a69067c0b12783fac6303fa85e8eb34d174add67ce3674382.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
108,
81,
504,
129.0
],
"spans": [],
"index": 0
},
{
"bbox": [
108,
129.0,
504,
177.0
],
"spans": [],
"index": 1
},
{
"bbox": [
108,
177.0,
504,
225.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
237,
505,
271
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
237,
505,
250
],
"spans": [
{
"bbox": [
105,
237,
378,
250
],
"score": 1.0,
"content": "Figure 4: Performance metrics for graph CNE compared to graph",
"type": "text"
},
{
"bbox": [
378,
239,
383,
248
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
383,
237,
409,
250
],
"score": 1.0,
"content": "-SNE:",
"type": "text"
},
{
"bbox": [
410,
237,
432,
248
],
"score": 0.48,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
432,
237,
460,
250
],
"score": 1.0,
"content": "recall,",
"type": "text"
},
{
"bbox": [
460,
237,
482,
248
],
"score": 0.63,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
483,
237,
505,
250
],
"score": 1.0,
"content": "clas-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
249,
505,
261
],
"spans": [
{
"bbox": [
106,
249,
505,
261
],
"score": 1.0,
"content": "sification accuracy, and linear accuracy. Shading shows standard deviation over five CNE runs.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
259,
312,
272
],
"spans": [
{
"bbox": [
106,
259,
312,
272
],
"score": 1.0,
"content": "Datasets are ordered by the increasing sample size.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
297,
505,
331
],
"lines": [
{
"bbox": [
105,
295,
505,
311
],
"spans": [
{
"bbox": [
105,
295,
289,
311
],
"score": 1.0,
"content": "Visually, the embeddings produced by graph",
"type": "text"
},
{
"bbox": [
289,
299,
294,
308
],
"score": 0.64,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
295,
295,
505,
311
],
"score": 1.0,
"content": "-SNE looked similar to DRGraph embeddings (Fig-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
308,
505,
321
],
"spans": [
{
"bbox": [
105,
308,
457,
321
],
"score": 1.0,
"content": "ures 2 and A.3), but showed richer within-class structure, in agreement with the higher",
"type": "text"
},
{
"bbox": [
458,
309,
479,
319
],
"score": 0.59,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
480,
308,
505,
321
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
318,
138,
333
],
"spans": [
{
"bbox": [
105,
318,
138,
333
],
"score": 1.0,
"content": "values.",
"type": "text"
}
],
"index": 8
}
],
"index": 7
},
{
"type": "text",
"bbox": [
107,
336,
505,
405
],
"lines": [
{
"bbox": [
106,
336,
504,
349
],
"spans": [
{
"bbox": [
106,
336,
504,
349
],
"score": 1.0,
"content": "We have also experimented with an alternative way to convert the adjacency matrix into the affinity",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
104,
345,
506,
362
],
"spans": [
{
"bbox": [
104,
345,
327,
362
],
"score": 1.0,
"content": "matrix: namely, to divide A by the sum of its elements:",
"type": "text"
},
{
"bbox": [
327,
347,
399,
361
],
"score": 0.92,
"content": "\\begin{array} { r } { \\mathbf { P } = \\mathbf { A } / \\sum _ { i j } \\hat { A _ { i j } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
399,
345,
506,
362
],
"score": 1.0,
"content": ". This approach resulted in",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
359,
505,
372
],
"spans": [
{
"bbox": [
105,
359,
137,
372
],
"score": 1.0,
"content": "similar",
"type": "text"
},
{
"bbox": [
137,
360,
159,
370
],
"score": 0.68,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
159,
359,
202,
372
],
"score": 1.0,
"content": "recall and",
"type": "text"
},
{
"bbox": [
202,
360,
224,
370
],
"score": 0.73,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
224,
359,
505,
372
],
"score": 1.0,
"content": "accuracy values, but gave visually unpleasing embeddings, with low-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
371,
505,
384
],
"spans": [
{
"bbox": [
105,
371,
505,
384
],
"score": 1.0,
"content": "degree nodes pushed out to the periphery (Figure A.2). Furthermore, we experimented with various",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
381,
506,
395
],
"spans": [
{
"bbox": [
105,
381,
506,
395
],
"score": 1.0,
"content": "initialization schemes, but found that on our graphs, random initialization performed very similar to",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
393,
295,
405
],
"spans": [
{
"bbox": [
105,
393,
295,
405
],
"score": 1.0,
"content": "the default Laplacian Eigenmaps initialization.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5
},
{
"type": "title",
"bbox": [
108,
428,
453,
440
],
"lines": [
{
"bbox": [
105,
428,
455,
441
],
"spans": [
{
"bbox": [
105,
428,
455,
441
],
"score": 1.0,
"content": "6 NODE-LEVEL GRAPH CONTRASTIVE LEARNING VIA GRAPH CNE",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "text",
"bbox": [
107,
456,
505,
501
],
"lines": [
{
"bbox": [
105,
456,
505,
469
],
"spans": [
{
"bbox": [
105,
456,
149,
469
],
"score": 1.0,
"content": "Similar to",
"type": "text"
},
{
"bbox": [
150,
457,
155,
466
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
155,
456,
429,
469
],
"score": 1.0,
"content": "-SNE, the CNE algorithm consists of two steps. First, it builds the",
"type": "text"
},
{
"bbox": [
430,
457,
452,
467
],
"score": 0.59,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
452,
456,
505,
469
],
"score": 1.0,
"content": "graph of the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
467,
505,
480
],
"spans": [
{
"bbox": [
105,
467,
505,
480
],
"score": 1.0,
"content": "data. Second, it optimizes the embedding (in our case, parametric embedding) using a contrastive",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
478,
505,
491
],
"spans": [
{
"bbox": [
105,
478,
505,
491
],
"score": 1.0,
"content": "loss function such as InfoNCE to make neighbors be close in the embedding (Section 3.2). For graph",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
488,
374,
502
],
"spans": [
{
"bbox": [
105,
488,
374,
502
],
"score": 1.0,
"content": "CNE, we omit the first step and provide the graph to CNE directly.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5
},
{
"type": "text",
"bbox": [
106,
506,
505,
627
],
"lines": [
{
"bbox": [
106,
506,
506,
519
],
"spans": [
{
"bbox": [
106,
506,
398,
519
],
"score": 1.0,
"content": "We used parametric CNE models, setting the output dimensionality to",
"type": "text"
},
{
"bbox": [
399,
506,
428,
516
],
"score": 0.89,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
428,
506,
448,
519
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
448,
506,
487,
517
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
487,
506,
506,
519
],
"score": 1.0,
"content": ". In",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
518,
505,
529
],
"spans": [
{
"bbox": [
106,
518,
505,
529
],
"score": 1.0,
"content": "both cases we used a fully-connected network (MLP), as is default in CNE, with the number of",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
104,
528,
505,
540
],
"spans": [
{
"bbox": [
104,
528,
199,
540
],
"score": 1.0,
"content": "neurons in each layer",
"type": "text"
},
{
"bbox": [
199,
528,
340,
539
],
"score": 0.92,
"content": "D \\ \\to \\ 1 0 0 \\ \\to \\ 1 0 0 \\ \\to \\ 1 0 0 \\ \\to \\ d",
"type": "inline_equation"
},
{
"bbox": [
341,
528,
374,
540
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
374,
529,
384,
538
],
"score": 0.8,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
384,
528,
505,
540
],
"score": 1.0,
"content": "is the number of input node",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
540,
504,
551
],
"spans": [
{
"bbox": [
106,
540,
504,
551
],
"score": 1.0,
"content": "features (Table 1). For both dimensionalities we used the InfoNCE loss. Following Damrich et al.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
549,
505,
563
],
"spans": [
{
"bbox": [
105,
549,
405,
563
],
"score": 1.0,
"content": "(2022), we used the cosine distance and the Gaussian similarity kernel for",
"type": "text"
},
{
"bbox": [
405,
550,
441,
560
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
441,
549,
505,
563
],
"score": 1.0,
"content": ", mimicking the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
561,
505,
574
],
"spans": [
{
"bbox": [
105,
561,
505,
574
],
"score": 1.0,
"content": "standard SimCLR setup (Chen et al., 2020), and the Euclidean distance and the Cauchy similarity",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
149,
585
],
"score": 1.0,
"content": "kernel for",
"type": "text"
},
{
"bbox": [
149,
572,
175,
582
],
"score": 0.89,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
176,
572,
277,
585
],
"score": 1.0,
"content": ", mimicking the standard",
"type": "text"
},
{
"bbox": [
277,
573,
282,
582
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
282,
572,
506,
585
],
"score": 1.0,
"content": "-SNE setup. We set the number of negative samples to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
582,
505,
596
],
"spans": [
{
"bbox": [
106,
582,
410,
596
],
"score": 1.0,
"content": "100 (increasing it from the default 5 improved the results), and batch size to",
"type": "text"
},
{
"bbox": [
411,
583,
490,
595
],
"score": 0.84,
"content": "\\operatorname* { m i n } \\{ \\bar { 1 } 0 2 4 , | \\mathcal { V } | / \\bar { 1 } 0 \\}",
"type": "inline_equation"
},
{
"bbox": [
490,
582,
505,
596
],
"score": 1.0,
"content": "(in",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
594,
505,
607
],
"spans": [
{
"bbox": [
105,
594,
505,
607
],
"score": 1.0,
"content": "pilot experiments we noticed that small graphs required smaller batch sizes for good convergence).",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
506,
618
],
"score": 1.0,
"content": "The number of epochs was set to 100. Optimization parameters were left at default values: Adam",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
616,
338,
628
],
"spans": [
{
"bbox": [
105,
616,
338,
628
],
"score": 1.0,
"content": "optimizer (Kingma & Ba, 2015) with learning rate 0.001.",
"type": "text"
}
],
"index": 30
}
],
"index": 25
},
{
"type": "text",
"bbox": [
106,
632,
505,
732
],
"lines": [
{
"bbox": [
106,
633,
505,
645
],
"spans": [
{
"bbox": [
106,
633,
189,
645
],
"score": 1.0,
"content": "Compared to graph",
"type": "text"
},
{
"bbox": [
190,
634,
195,
643
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
195,
633,
317,
645
],
"score": 1.0,
"content": "-SNE, graph CNE, with both",
"type": "text"
},
{
"bbox": [
317,
633,
347,
643
],
"score": 0.9,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
348,
633,
367,
645
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
367,
633,
407,
644
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
407,
633,
456,
645
],
"score": 1.0,
"content": ", had lower",
"type": "text"
},
{
"bbox": [
457,
633,
478,
643
],
"score": 0.77,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
479,
633,
505,
645
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
643,
505,
656
],
"spans": [
{
"bbox": [
106,
643,
458,
656
],
"score": 1.0,
"content": "(Figure 4). This is likely because graph CNE had to use node features, whereas graph",
"type": "text"
},
{
"bbox": [
459,
645,
463,
654
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
464,
643,
505,
656
],
"score": 1.0,
"content": "-SNE was",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
655,
506,
668
],
"spans": [
{
"bbox": [
105,
655,
506,
668
],
"score": 1.0,
"content": "unconstrained by them and optimized graph neighborhood preservation directly. At the same time,",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
664,
505,
679
],
"spans": [
{
"bbox": [
106,
666,
128,
676
],
"score": 0.56,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
128,
664,
505,
679
],
"score": 1.0,
"content": "accuracy was very similar (Figure 4) on all datasets, apart from the ARX dataset. The compar-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
677,
505,
690
],
"spans": [
{
"bbox": [
105,
677,
505,
690
],
"score": 1.0,
"content": "atively poor performance of graph CNE on the ARX dataset was likely due to ARX feature space",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
329,
700
],
"score": 1.0,
"content": "showing weak class separation (Table 2); whereas graph",
"type": "text"
},
{
"bbox": [
330,
689,
334,
698
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
335,
688,
505,
700
],
"score": 1.0,
"content": "-SNE does not use node features and hence",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
698,
505,
711
],
"spans": [
{
"bbox": [
105,
698,
505,
711
],
"score": 1.0,
"content": "is not influenced by the feature quality. Visually, two-dimensional graph CNE embeddings looked",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
194,
723
],
"score": 1.0,
"content": "very similar to graph",
"type": "text"
},
{
"bbox": [
195,
711,
199,
720
],
"score": 0.55,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
200,
709,
506,
723
],
"score": 1.0,
"content": "-SNE embeddings (Figure 1), even though the former were parametric and",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
721,
231,
733
],
"spans": [
{
"bbox": [
106,
721,
231,
733
],
"score": 1.0,
"content": "the latter were non-parametric.",
"type": "text"
}
],
"index": 39
}
],
"index": 35
}
],
"page_idx": 6,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
303,
751,
309,
759
],
"lines": [
{
"bbox": [
302,
750,
309,
762
],
"spans": [
{
"bbox": [
302,
750,
309,
762
],
"score": 1.0,
"content": "7",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
108,
81,
504,
225
],
"blocks": [
{
"type": "image_body",
"bbox": [
108,
81,
504,
225
],
"group_id": 0,
"lines": [
{
"bbox": [
108,
81,
504,
225
],
"spans": [
{
"bbox": [
108,
81,
504,
225
],
"score": 0.968,
"type": "image",
"image_path": "472c51903edfd06a69067c0b12783fac6303fa85e8eb34d174add67ce3674382.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
108,
81,
504,
129.0
],
"spans": [],
"index": 0
},
{
"bbox": [
108,
129.0,
504,
177.0
],
"spans": [],
"index": 1
},
{
"bbox": [
108,
177.0,
504,
225.0
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
237,
505,
271
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
237,
505,
250
],
"spans": [
{
"bbox": [
105,
237,
378,
250
],
"score": 1.0,
"content": "Figure 4: Performance metrics for graph CNE compared to graph",
"type": "text"
},
{
"bbox": [
378,
239,
383,
248
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
383,
237,
409,
250
],
"score": 1.0,
"content": "-SNE:",
"type": "text"
},
{
"bbox": [
410,
237,
432,
248
],
"score": 0.48,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
432,
237,
460,
250
],
"score": 1.0,
"content": "recall,",
"type": "text"
},
{
"bbox": [
460,
237,
482,
248
],
"score": 0.63,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
483,
237,
505,
250
],
"score": 1.0,
"content": "clas-",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
249,
505,
261
],
"spans": [
{
"bbox": [
106,
249,
505,
261
],
"score": 1.0,
"content": "sification accuracy, and linear accuracy. Shading shows standard deviation over five CNE runs.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
259,
312,
272
],
"spans": [
{
"bbox": [
106,
259,
312,
272
],
"score": 1.0,
"content": "Datasets are ordered by the increasing sample size.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
},
{
"type": "text",
"bbox": [
107,
297,
505,
331
],
"lines": [
{
"bbox": [
105,
295,
505,
311
],
"spans": [
{
"bbox": [
105,
295,
289,
311
],
"score": 1.0,
"content": "Visually, the embeddings produced by graph",
"type": "text"
},
{
"bbox": [
289,
299,
294,
308
],
"score": 0.64,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
295,
295,
505,
311
],
"score": 1.0,
"content": "-SNE looked similar to DRGraph embeddings (Fig-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
308,
505,
321
],
"spans": [
{
"bbox": [
105,
308,
457,
321
],
"score": 1.0,
"content": "ures 2 and A.3), but showed richer within-class structure, in agreement with the higher",
"type": "text"
},
{
"bbox": [
458,
309,
479,
319
],
"score": 0.59,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
480,
308,
505,
321
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
318,
138,
333
],
"spans": [
{
"bbox": [
105,
318,
138,
333
],
"score": 1.0,
"content": "values.",
"type": "text"
}
],
"index": 8
}
],
"index": 7,
"bbox_fs": [
105,
295,
505,
333
]
},
{
"type": "text",
"bbox": [
107,
336,
505,
405
],
"lines": [
{
"bbox": [
106,
336,
504,
349
],
"spans": [
{
"bbox": [
106,
336,
504,
349
],
"score": 1.0,
"content": "We have also experimented with an alternative way to convert the adjacency matrix into the affinity",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
104,
345,
506,
362
],
"spans": [
{
"bbox": [
104,
345,
327,
362
],
"score": 1.0,
"content": "matrix: namely, to divide A by the sum of its elements:",
"type": "text"
},
{
"bbox": [
327,
347,
399,
361
],
"score": 0.92,
"content": "\\begin{array} { r } { \\mathbf { P } = \\mathbf { A } / \\sum _ { i j } \\hat { A _ { i j } } } \\end{array}",
"type": "inline_equation"
},
{
"bbox": [
399,
345,
506,
362
],
"score": 1.0,
"content": ". This approach resulted in",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
359,
505,
372
],
"spans": [
{
"bbox": [
105,
359,
137,
372
],
"score": 1.0,
"content": "similar",
"type": "text"
},
{
"bbox": [
137,
360,
159,
370
],
"score": 0.68,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
159,
359,
202,
372
],
"score": 1.0,
"content": "recall and",
"type": "text"
},
{
"bbox": [
202,
360,
224,
370
],
"score": 0.73,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
224,
359,
505,
372
],
"score": 1.0,
"content": "accuracy values, but gave visually unpleasing embeddings, with low-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
371,
505,
384
],
"spans": [
{
"bbox": [
105,
371,
505,
384
],
"score": 1.0,
"content": "degree nodes pushed out to the periphery (Figure A.2). Furthermore, we experimented with various",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
381,
506,
395
],
"spans": [
{
"bbox": [
105,
381,
506,
395
],
"score": 1.0,
"content": "initialization schemes, but found that on our graphs, random initialization performed very similar to",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
393,
295,
405
],
"spans": [
{
"bbox": [
105,
393,
295,
405
],
"score": 1.0,
"content": "the default Laplacian Eigenmaps initialization.",
"type": "text"
}
],
"index": 14
}
],
"index": 11.5,
"bbox_fs": [
104,
336,
506,
405
]
},
{
"type": "title",
"bbox": [
108,
428,
453,
440
],
"lines": [
{
"bbox": [
105,
428,
455,
441
],
"spans": [
{
"bbox": [
105,
428,
455,
441
],
"score": 1.0,
"content": "6 NODE-LEVEL GRAPH CONTRASTIVE LEARNING VIA GRAPH CNE",
"type": "text"
}
],
"index": 15
}
],
"index": 15
},
{
"type": "text",
"bbox": [
107,
456,
505,
501
],
"lines": [
{
"bbox": [
105,
456,
505,
469
],
"spans": [
{
"bbox": [
105,
456,
149,
469
],
"score": 1.0,
"content": "Similar to",
"type": "text"
},
{
"bbox": [
150,
457,
155,
466
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
155,
456,
429,
469
],
"score": 1.0,
"content": "-SNE, the CNE algorithm consists of two steps. First, it builds the",
"type": "text"
},
{
"bbox": [
430,
457,
452,
467
],
"score": 0.59,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
452,
456,
505,
469
],
"score": 1.0,
"content": "graph of the",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
467,
505,
480
],
"spans": [
{
"bbox": [
105,
467,
505,
480
],
"score": 1.0,
"content": "data. Second, it optimizes the embedding (in our case, parametric embedding) using a contrastive",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
478,
505,
491
],
"spans": [
{
"bbox": [
105,
478,
505,
491
],
"score": 1.0,
"content": "loss function such as InfoNCE to make neighbors be close in the embedding (Section 3.2). For graph",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
488,
374,
502
],
"spans": [
{
"bbox": [
105,
488,
374,
502
],
"score": 1.0,
"content": "CNE, we omit the first step and provide the graph to CNE directly.",
"type": "text"
}
],
"index": 19
}
],
"index": 17.5,
"bbox_fs": [
105,
456,
505,
502
]
},
{
"type": "text",
"bbox": [
106,
506,
505,
627
],
"lines": [
{
"bbox": [
106,
506,
506,
519
],
"spans": [
{
"bbox": [
106,
506,
398,
519
],
"score": 1.0,
"content": "We used parametric CNE models, setting the output dimensionality to",
"type": "text"
},
{
"bbox": [
399,
506,
428,
516
],
"score": 0.89,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
428,
506,
448,
519
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
448,
506,
487,
517
],
"score": 0.89,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
487,
506,
506,
519
],
"score": 1.0,
"content": ". In",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
106,
518,
505,
529
],
"spans": [
{
"bbox": [
106,
518,
505,
529
],
"score": 1.0,
"content": "both cases we used a fully-connected network (MLP), as is default in CNE, with the number of",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
104,
528,
505,
540
],
"spans": [
{
"bbox": [
104,
528,
199,
540
],
"score": 1.0,
"content": "neurons in each layer",
"type": "text"
},
{
"bbox": [
199,
528,
340,
539
],
"score": 0.92,
"content": "D \\ \\to \\ 1 0 0 \\ \\to \\ 1 0 0 \\ \\to \\ 1 0 0 \\ \\to \\ d",
"type": "inline_equation"
},
{
"bbox": [
341,
528,
374,
540
],
"score": 1.0,
"content": ", where",
"type": "text"
},
{
"bbox": [
374,
529,
384,
538
],
"score": 0.8,
"content": "D",
"type": "inline_equation"
},
{
"bbox": [
384,
528,
505,
540
],
"score": 1.0,
"content": "is the number of input node",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
540,
504,
551
],
"spans": [
{
"bbox": [
106,
540,
504,
551
],
"score": 1.0,
"content": "features (Table 1). For both dimensionalities we used the InfoNCE loss. Following Damrich et al.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
549,
505,
563
],
"spans": [
{
"bbox": [
105,
549,
405,
563
],
"score": 1.0,
"content": "(2022), we used the cosine distance and the Gaussian similarity kernel for",
"type": "text"
},
{
"bbox": [
405,
550,
441,
560
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
441,
549,
505,
563
],
"score": 1.0,
"content": ", mimicking the",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
561,
505,
574
],
"spans": [
{
"bbox": [
105,
561,
505,
574
],
"score": 1.0,
"content": "standard SimCLR setup (Chen et al., 2020), and the Euclidean distance and the Cauchy similarity",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
572,
506,
585
],
"spans": [
{
"bbox": [
105,
572,
149,
585
],
"score": 1.0,
"content": "kernel for",
"type": "text"
},
{
"bbox": [
149,
572,
175,
582
],
"score": 0.89,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
176,
572,
277,
585
],
"score": 1.0,
"content": ", mimicking the standard",
"type": "text"
},
{
"bbox": [
277,
573,
282,
582
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
282,
572,
506,
585
],
"score": 1.0,
"content": "-SNE setup. We set the number of negative samples to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
582,
505,
596
],
"spans": [
{
"bbox": [
106,
582,
410,
596
],
"score": 1.0,
"content": "100 (increasing it from the default 5 improved the results), and batch size to",
"type": "text"
},
{
"bbox": [
411,
583,
490,
595
],
"score": 0.84,
"content": "\\operatorname* { m i n } \\{ \\bar { 1 } 0 2 4 , | \\mathcal { V } | / \\bar { 1 } 0 \\}",
"type": "inline_equation"
},
{
"bbox": [
490,
582,
505,
596
],
"score": 1.0,
"content": "(in",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
594,
505,
607
],
"spans": [
{
"bbox": [
105,
594,
505,
607
],
"score": 1.0,
"content": "pilot experiments we noticed that small graphs required smaller batch sizes for good convergence).",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
105,
604,
506,
618
],
"spans": [
{
"bbox": [
105,
604,
506,
618
],
"score": 1.0,
"content": "The number of epochs was set to 100. Optimization parameters were left at default values: Adam",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
616,
338,
628
],
"spans": [
{
"bbox": [
105,
616,
338,
628
],
"score": 1.0,
"content": "optimizer (Kingma & Ba, 2015) with learning rate 0.001.",
"type": "text"
}
],
"index": 30
}
],
"index": 25,
"bbox_fs": [
104,
506,
506,
628
]
},
{
"type": "text",
"bbox": [
106,
632,
505,
732
],
"lines": [
{
"bbox": [
106,
633,
505,
645
],
"spans": [
{
"bbox": [
106,
633,
189,
645
],
"score": 1.0,
"content": "Compared to graph",
"type": "text"
},
{
"bbox": [
190,
634,
195,
643
],
"score": 0.66,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
195,
633,
317,
645
],
"score": 1.0,
"content": "-SNE, graph CNE, with both",
"type": "text"
},
{
"bbox": [
317,
633,
347,
643
],
"score": 0.9,
"content": "d = 2",
"type": "inline_equation"
},
{
"bbox": [
348,
633,
367,
645
],
"score": 1.0,
"content": "and",
"type": "text"
},
{
"bbox": [
367,
633,
407,
644
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
407,
633,
456,
645
],
"score": 1.0,
"content": ", had lower",
"type": "text"
},
{
"bbox": [
457,
633,
478,
643
],
"score": 0.77,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
479,
633,
505,
645
],
"score": 1.0,
"content": "recall",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
106,
643,
505,
656
],
"spans": [
{
"bbox": [
106,
643,
458,
656
],
"score": 1.0,
"content": "(Figure 4). This is likely because graph CNE had to use node features, whereas graph",
"type": "text"
},
{
"bbox": [
459,
645,
463,
654
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
464,
643,
505,
656
],
"score": 1.0,
"content": "-SNE was",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
105,
655,
506,
668
],
"spans": [
{
"bbox": [
105,
655,
506,
668
],
"score": 1.0,
"content": "unconstrained by them and optimized graph neighborhood preservation directly. At the same time,",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
106,
664,
505,
679
],
"spans": [
{
"bbox": [
106,
666,
128,
676
],
"score": 0.56,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
128,
664,
505,
679
],
"score": 1.0,
"content": "accuracy was very similar (Figure 4) on all datasets, apart from the ARX dataset. The compar-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
677,
505,
690
],
"spans": [
{
"bbox": [
105,
677,
505,
690
],
"score": 1.0,
"content": "atively poor performance of graph CNE on the ARX dataset was likely due to ARX feature space",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
106,
688,
505,
700
],
"spans": [
{
"bbox": [
106,
688,
329,
700
],
"score": 1.0,
"content": "showing weak class separation (Table 2); whereas graph",
"type": "text"
},
{
"bbox": [
330,
689,
334,
698
],
"score": 0.67,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
335,
688,
505,
700
],
"score": 1.0,
"content": "-SNE does not use node features and hence",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
698,
505,
711
],
"spans": [
{
"bbox": [
105,
698,
505,
711
],
"score": 1.0,
"content": "is not influenced by the feature quality. Visually, two-dimensional graph CNE embeddings looked",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
709,
506,
723
],
"spans": [
{
"bbox": [
105,
709,
194,
723
],
"score": 1.0,
"content": "very similar to graph",
"type": "text"
},
{
"bbox": [
195,
711,
199,
720
],
"score": 0.55,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
200,
709,
506,
723
],
"score": 1.0,
"content": "-SNE embeddings (Figure 1), even though the former were parametric and",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
106,
721,
231,
733
],
"spans": [
{
"bbox": [
106,
721,
231,
733
],
"score": 1.0,
"content": "the latter were non-parametric.",
"type": "text"
}
],
"index": 39
}
],
"index": 35,
"bbox_fs": [
105,
633,
506,
733
]
}
]
},
{
"preproc_blocks": [
{
"type": "table",
"bbox": [
106,
168,
506,
371
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
89,
505,
167
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
88,
505,
103
],
"spans": [
{
"bbox": [
105,
88,
276,
103
],
"score": 1.0,
"content": "Table 2: Linear classification accuracy (in",
"type": "text"
},
{
"bbox": [
276,
90,
286,
100
],
"score": 0.58,
"content": "\\%",
"type": "inline_equation"
},
{
"bbox": [
286,
88,
505,
103
],
"score": 1.0,
"content": ") of graph CNE and existing graph contrastive learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
100,
505,
113
],
"spans": [
{
"bbox": [
106,
100,
469,
113
],
"score": 1.0,
"content": "algorithms. Output dimensionality of CNE is indicated in brackets. The line marked by",
"type": "text"
},
{
"bbox": [
469,
102,
477,
110
],
"score": 0.74,
"content": "\\star",
"type": "inline_equation"
},
{
"bbox": [
477,
100,
505,
113
],
"score": 1.0,
"content": "shows",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
110,
506,
124
],
"spans": [
{
"bbox": [
106,
111,
128,
122
],
"score": 0.36,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
129,
110,
371,
124
],
"score": 1.0,
"content": "accuracy instead of linear accuracy. CNE values are mean",
"type": "text"
},
{
"bbox": [
371,
112,
381,
122
],
"score": 0.62,
"content": "\\pm",
"type": "inline_equation"
},
{
"bbox": [
382,
110,
506,
124
],
"score": 1.0,
"content": "standard deviation across five",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
122,
505,
135
],
"spans": [
{
"bbox": [
105,
122,
505,
135
],
"score": 1.0,
"content": "training runs. Non-CNE values are taken from Zhang et al. (2022), MLP values are taken from",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
134,
504,
145
],
"spans": [
{
"bbox": [
106,
134,
292,
145
],
"score": 1.0,
"content": "https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
292,
135,
299,
143
],
"score": 0.43,
"content": "{ . } = { }",
"type": "inline_equation"
},
{
"bbox": [
300,
134,
399,
145
],
"score": 1.0,
"content": "dSYkYNNZkV¬eId",
"type": "text"
},
{
"bbox": [
400,
135,
407,
143
],
"score": 0.41,
"content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =",
"type": "inline_equation"
},
{
"bbox": [
407,
134,
504,
145
],
"score": 1.0,
"content": "aLQzIXVy0w and Guo",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
143,
505,
157
],
"spans": [
{
"bbox": [
105,
143,
505,
157
],
"score": 1.0,
"content": "et al. (2023). OOM denotes out-of-memory error. Datasets are ordered by the increasing sample",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
155,
417,
168
],
"spans": [
{
"bbox": [
105,
155,
417,
168
],
"score": 1.0,
"content": "size. For comparison, the first row shows linear accuracy in the feature space.",
"type": "text"
}
],
"index": 6
}
],
"index": 3
},
{
"type": "table_body",
"bbox": [
106,
168,
506,
371
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
168,
506,
371
],
"spans": [
{
"bbox": [
106,
168,
506,
371
],
"score": 0.985,
"html": " | CSR | COR | APH | ACO | PUB | ARX |
| Feature space | 70.3 | 68.6 | 90.7 | 79.6 | 87.8 | 55.1 |
| Graph CNE (2) | 65.4 ± 2.2 | 62.7 ± 6.2 | 73.2 ± 1.4 | 77.1 ±0.7 | 66.9 ± 2.3 | 41.7 ± 0.8 |
| Graph CNE (2) * | 72.1 ± 1.5 | 78.1± 3.2 | 92.9 ± 0.3 | 89.0± 0.2 | 77.2 ±0.6 | 45.3 ± 0.2 |
| Graph CNE (128) | 72.0 ± 1.3 | 80.0± 1.2 | 92.9 ± 0.5 | 86.8±0.7 | 84.6±0.6 | 52.9 ± 0.3 |
| GRACE | 71.2 ± 0.5 | 81.9 ± 0.4 | 92.2± 0.2 | 86.3± 0.3 | 80.6±0.4 | 0OM |
| GCA | 72.1± 0.4 | 82.3 ± 0.4 | 92.5 ± 0.1 | 87.9 ± 0.3 | 80.7± 0.5 | OOM |
| MVGRL | 73.3 ± 0.5 | 83.5 ± 0.4 | 91.7± 0.1 | 87.5 ± 0.1 | 80.1±0.7 | 0OM |
| DGI | 71.8±0.7 | 82.3 ± 0.6 | 91.6 ± 0.2 | 83.9± 0.5 | 76.8± 0.6 | 71.2 ± 0.2 |
| BGRL | 71.1±0.8 | 82.7 ± 0.6 | 93.1 ± 0.3 | 89.7 ± 0.4 | 79.6± 0.5 | 72.7± 0.2 |
| CCA-SSG | 73.1 ± 0.3 | 84.2 ± 0.4 | 93.1 ± 0.1 | 88.7± 0.3 | 81.6± 0.4 | 72.3 ± 0.2 |
| AF-GCL | 72.0±0.4 | 83.2± 0.2 | 92.5 ± 0.3 | 89.7±0.2 | 79.1 ± 0.8 | |
| AFGRL | 68.7± 0.3 | 81.3± 0.2 | 93.2 ± 0.3 | 89.9 ± 0.3 | 80.6± 0.4 | 0OM |
| Local-GCL | 73.6 ± 0.4 | 84.5± 0.4 | 93.3 ± 0.4 | 88.8±0.4 | 82.1 ± 0.5 | 71.3 ± 0.3 |
| Local-GCL,MLP | 70.3± 0.6 | 78.3± 0.5 | 90.9±0.4 | 82.4±0.5 | 79.6± 0.5 | |
| GRACE, MLP | 65.5 ± 2.6 | 67.7 ± 0.9 | 87.9 ± 0.6 | 80.9 ± 1.2 | 83.3± 0.5 | |
",
"type": "table",
"image_path": "e72560a8e6b78995d863b84f76cc6f8e4908c62d027c3ff94e9094eea925efd4.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
106,
168,
506,
235.66666666666669
],
"spans": [],
"index": 7
},
{
"bbox": [
106,
235.66666666666669,
506,
303.33333333333337
],
"spans": [],
"index": 8
},
{
"bbox": [
106,
303.33333333333337,
506,
371.00000000000006
],
"spans": [],
"index": 9
}
]
}
],
"index": 5.5
},
{
"type": "text",
"bbox": [
107,
394,
505,
450
],
"lines": [
{
"bbox": [
106,
394,
505,
408
],
"spans": [
{
"bbox": [
106,
394,
205,
408
],
"score": 1.0,
"content": "As expected, CNE with",
"type": "text"
},
{
"bbox": [
206,
395,
243,
406
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
243,
394,
505,
408
],
"score": 1.0,
"content": ", yielded considerably higher linear classification accuracy com-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
406,
505,
419
],
"spans": [
{
"bbox": [
105,
406,
505,
419
],
"score": 1.0,
"content": "pared to both 2-dimensional embeddings (Figure 4). In terms of linear accuracy, graph CNE per-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
416,
505,
430
],
"spans": [
{
"bbox": [
105,
416,
505,
430
],
"score": 1.0,
"content": "formed comparably to the state-of-the-art graph contrastive learning (GCL) algorithms1 (Table 2).",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
428,
505,
440
],
"spans": [
{
"bbox": [
105,
428,
505,
440
],
"score": 1.0,
"content": "Graph CNE achieved the best results on one of the datasets (PUB), and had close to the best results",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
439,
263,
451
],
"spans": [
{
"bbox": [
106,
439,
263,
451
],
"score": 1.0,
"content": "on other datasets, apart from the ARX.",
"type": "text"
}
],
"index": 14
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
455,
505,
576
],
"lines": [
{
"bbox": [
105,
456,
505,
468
],
"spans": [
{
"bbox": [
105,
456,
505,
468
],
"score": 1.0,
"content": "Note that graph CNE was at disadvantage compared to all other GCL methods listed in Table 2",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
467,
505,
479
],
"spans": [
{
"bbox": [
105,
467,
505,
479
],
"score": 1.0,
"content": "because it used an MLP network, whereas other GCL methods traditionally use graph convolutional",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
477,
505,
491
],
"spans": [
{
"bbox": [
105,
477,
505,
491
],
"score": 1.0,
"content": "networks (GCN). GCN takes the entire graph as input and uses message passing, which pulls to-",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
488,
506,
502
],
"spans": [
{
"bbox": [
104,
488,
506,
502
],
"score": 1.0,
"content": "gether embeddings of connected nodes and helps to obtain better embeddings. However, GCN is",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
499,
505,
512
],
"spans": [
{
"bbox": [
105,
499,
505,
512
],
"score": 1.0,
"content": "not able to transform one node at a time, and so a trained GCN cannot be applied to a new, held-out",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
510,
505,
523
],
"spans": [
{
"bbox": [
105,
510,
505,
523
],
"score": 1.0,
"content": "node. In contrast, our graph CNE with MLP can (after training) process one node at a time, which",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
522,
505,
534
],
"spans": [
{
"bbox": [
105,
522,
505,
534
],
"score": 1.0,
"content": "we consider more appropriate for node-level graph learning (see Discussion). There are very few",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
532,
505,
544
],
"spans": [
{
"bbox": [
105,
532,
505,
544
],
"score": 1.0,
"content": "GCL results based on the MLP architecture reported in the literature. Two examples are Local-GCL",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
543,
504,
556
],
"spans": [
{
"bbox": [
105,
543,
504,
556
],
"score": 1.0,
"content": "and GRACE trained with MLP architecture (reported in the OpenReview discussion of Zhang et al.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
554,
505,
568
],
"spans": [
{
"bbox": [
105,
554,
505,
568
],
"score": 1.0,
"content": "(2022) and in Guo et al. (2023) respectively, Table 2): both had lower accuracy compared to our",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
565,
217,
578
],
"spans": [
{
"bbox": [
105,
565,
217,
578
],
"score": 1.0,
"content": "graph CNE on all datasets.",
"type": "text"
}
],
"index": 25
}
],
"index": 20
},
{
"type": "text",
"bbox": [
108,
582,
505,
615
],
"lines": [
{
"bbox": [
105,
581,
505,
595
],
"spans": [
{
"bbox": [
105,
581,
505,
595
],
"score": 1.0,
"content": "For the ARX graph, we did not find any existing MLP-based results. Lower performance of graph",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
593,
505,
606
],
"spans": [
{
"bbox": [
106,
593,
505,
606
],
"score": 1.0,
"content": "CNE compared to GCN-based GCL methods was, again, likely due to the feature space of this graph",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
604,
331,
617
],
"spans": [
{
"bbox": [
105,
604,
331,
617
],
"score": 1.0,
"content": "showing only weak class separation (Table 2, first row).",
"type": "text"
}
],
"index": 28
}
],
"index": 27
},
{
"type": "title",
"bbox": [
108,
635,
190,
648
],
"lines": [
{
"bbox": [
104,
633,
192,
651
],
"spans": [
{
"bbox": [
104,
633,
192,
651
],
"score": 1.0,
"content": "7 DISCUSSION",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
662,
443,
673
],
"lines": [
{
"bbox": [
105,
660,
446,
676
],
"spans": [
{
"bbox": [
105,
660,
446,
676
],
"score": 1.0,
"content": "Summary Our paper makes three contributions, two practical and one conceptual:",
"type": "text"
}
],
"index": 30
}
],
"index": 30
},
{
"type": "text",
"bbox": [
114,
685,
504,
707
],
"lines": [
{
"bbox": [
115,
685,
505,
697
],
"spans": [
{
"bbox": [
115,
685,
345,
697
],
"score": 1.0,
"content": "i. We suggested a novel graph layout algorithm, graph",
"type": "text"
},
{
"bbox": [
345,
686,
350,
695
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
350,
685,
505,
697
],
"score": 1.0,
"content": "-SNE, and showed that it outperforms",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
127,
695,
352,
708
],
"spans": [
{
"bbox": [
127,
695,
352,
708
],
"score": 1.0,
"content": "existing competitors in preserving local graph structure.",
"type": "text"
}
],
"index": 32
}
],
"index": 31.5
}
],
"page_idx": 7,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
117,
721,
469,
732
],
"lines": [
{
"bbox": [
119,
720,
470,
733
],
"spans": [
{
"bbox": [
119,
720,
470,
733
],
"score": 1.0,
"content": "1We did not measure their performance ourselves, but took the values directly from the literature.",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
25,
308,
38
],
"spans": [
{
"bbox": [
107,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
760
],
"lines": [
{
"bbox": [
300,
750,
309,
761
],
"spans": [
{
"bbox": [
300,
750,
309,
761
],
"score": 1.0,
"content": "8",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "table",
"bbox": [
106,
168,
506,
371
],
"blocks": [
{
"type": "table_caption",
"bbox": [
106,
89,
505,
167
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
88,
505,
103
],
"spans": [
{
"bbox": [
105,
88,
276,
103
],
"score": 1.0,
"content": "Table 2: Linear classification accuracy (in",
"type": "text"
},
{
"bbox": [
276,
90,
286,
100
],
"score": 0.58,
"content": "\\%",
"type": "inline_equation"
},
{
"bbox": [
286,
88,
505,
103
],
"score": 1.0,
"content": ") of graph CNE and existing graph contrastive learning",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
106,
100,
505,
113
],
"spans": [
{
"bbox": [
106,
100,
469,
113
],
"score": 1.0,
"content": "algorithms. Output dimensionality of CNE is indicated in brackets. The line marked by",
"type": "text"
},
{
"bbox": [
469,
102,
477,
110
],
"score": 0.74,
"content": "\\star",
"type": "inline_equation"
},
{
"bbox": [
477,
100,
505,
113
],
"score": 1.0,
"content": "shows",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
106,
110,
506,
124
],
"spans": [
{
"bbox": [
106,
111,
128,
122
],
"score": 0.36,
"content": "k \\mathbf { N N }",
"type": "inline_equation"
},
{
"bbox": [
129,
110,
371,
124
],
"score": 1.0,
"content": "accuracy instead of linear accuracy. CNE values are mean",
"type": "text"
},
{
"bbox": [
371,
112,
381,
122
],
"score": 0.62,
"content": "\\pm",
"type": "inline_equation"
},
{
"bbox": [
382,
110,
506,
124
],
"score": 1.0,
"content": "standard deviation across five",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
122,
505,
135
],
"spans": [
{
"bbox": [
105,
122,
505,
135
],
"score": 1.0,
"content": "training runs. Non-CNE values are taken from Zhang et al. (2022), MLP values are taken from",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
134,
504,
145
],
"spans": [
{
"bbox": [
106,
134,
292,
145
],
"score": 1.0,
"content": "https://openreview.net/forum?id",
"type": "text"
},
{
"bbox": [
292,
135,
299,
143
],
"score": 0.43,
"content": "{ . } = { }",
"type": "inline_equation"
},
{
"bbox": [
300,
134,
399,
145
],
"score": 1.0,
"content": "dSYkYNNZkV¬eId",
"type": "text"
},
{
"bbox": [
400,
135,
407,
143
],
"score": 0.41,
"content": "\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =",
"type": "inline_equation"
},
{
"bbox": [
407,
134,
504,
145
],
"score": 1.0,
"content": "aLQzIXVy0w and Guo",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
143,
505,
157
],
"spans": [
{
"bbox": [
105,
143,
505,
157
],
"score": 1.0,
"content": "et al. (2023). OOM denotes out-of-memory error. Datasets are ordered by the increasing sample",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
155,
417,
168
],
"spans": [
{
"bbox": [
105,
155,
417,
168
],
"score": 1.0,
"content": "size. For comparison, the first row shows linear accuracy in the feature space.",
"type": "text"
}
],
"index": 6
}
],
"index": 3
},
{
"type": "table_body",
"bbox": [
106,
168,
506,
371
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
168,
506,
371
],
"spans": [
{
"bbox": [
106,
168,
506,
371
],
"score": 0.985,
"html": " | CSR | COR | APH | ACO | PUB | ARX |
| Feature space | 70.3 | 68.6 | 90.7 | 79.6 | 87.8 | 55.1 |
| Graph CNE (2) | 65.4 ± 2.2 | 62.7 ± 6.2 | 73.2 ± 1.4 | 77.1 ±0.7 | 66.9 ± 2.3 | 41.7 ± 0.8 |
| Graph CNE (2) * | 72.1 ± 1.5 | 78.1± 3.2 | 92.9 ± 0.3 | 89.0± 0.2 | 77.2 ±0.6 | 45.3 ± 0.2 |
| Graph CNE (128) | 72.0 ± 1.3 | 80.0± 1.2 | 92.9 ± 0.5 | 86.8±0.7 | 84.6±0.6 | 52.9 ± 0.3 |
| GRACE | 71.2 ± 0.5 | 81.9 ± 0.4 | 92.2± 0.2 | 86.3± 0.3 | 80.6±0.4 | 0OM |
| GCA | 72.1± 0.4 | 82.3 ± 0.4 | 92.5 ± 0.1 | 87.9 ± 0.3 | 80.7± 0.5 | OOM |
| MVGRL | 73.3 ± 0.5 | 83.5 ± 0.4 | 91.7± 0.1 | 87.5 ± 0.1 | 80.1±0.7 | 0OM |
| DGI | 71.8±0.7 | 82.3 ± 0.6 | 91.6 ± 0.2 | 83.9± 0.5 | 76.8± 0.6 | 71.2 ± 0.2 |
| BGRL | 71.1±0.8 | 82.7 ± 0.6 | 93.1 ± 0.3 | 89.7 ± 0.4 | 79.6± 0.5 | 72.7± 0.2 |
| CCA-SSG | 73.1 ± 0.3 | 84.2 ± 0.4 | 93.1 ± 0.1 | 88.7± 0.3 | 81.6± 0.4 | 72.3 ± 0.2 |
| AF-GCL | 72.0±0.4 | 83.2± 0.2 | 92.5 ± 0.3 | 89.7±0.2 | 79.1 ± 0.8 | |
| AFGRL | 68.7± 0.3 | 81.3± 0.2 | 93.2 ± 0.3 | 89.9 ± 0.3 | 80.6± 0.4 | 0OM |
| Local-GCL | 73.6 ± 0.4 | 84.5± 0.4 | 93.3 ± 0.4 | 88.8±0.4 | 82.1 ± 0.5 | 71.3 ± 0.3 |
| Local-GCL,MLP | 70.3± 0.6 | 78.3± 0.5 | 90.9±0.4 | 82.4±0.5 | 79.6± 0.5 | |
| GRACE, MLP | 65.5 ± 2.6 | 67.7 ± 0.9 | 87.9 ± 0.6 | 80.9 ± 1.2 | 83.3± 0.5 | |
",
"type": "table",
"image_path": "e72560a8e6b78995d863b84f76cc6f8e4908c62d027c3ff94e9094eea925efd4.jpg"
}
]
}
],
"index": 8,
"virtual_lines": [
{
"bbox": [
106,
168,
506,
235.66666666666669
],
"spans": [],
"index": 7
},
{
"bbox": [
106,
235.66666666666669,
506,
303.33333333333337
],
"spans": [],
"index": 8
},
{
"bbox": [
106,
303.33333333333337,
506,
371.00000000000006
],
"spans": [],
"index": 9
}
]
}
],
"index": 5.5
},
{
"type": "text",
"bbox": [
107,
394,
505,
450
],
"lines": [
{
"bbox": [
106,
394,
505,
408
],
"spans": [
{
"bbox": [
106,
394,
205,
408
],
"score": 1.0,
"content": "As expected, CNE with",
"type": "text"
},
{
"bbox": [
206,
395,
243,
406
],
"score": 0.9,
"content": "d = 1 2 8",
"type": "inline_equation"
},
{
"bbox": [
243,
394,
505,
408
],
"score": 1.0,
"content": ", yielded considerably higher linear classification accuracy com-",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
406,
505,
419
],
"spans": [
{
"bbox": [
105,
406,
505,
419
],
"score": 1.0,
"content": "pared to both 2-dimensional embeddings (Figure 4). In terms of linear accuracy, graph CNE per-",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
416,
505,
430
],
"spans": [
{
"bbox": [
105,
416,
505,
430
],
"score": 1.0,
"content": "formed comparably to the state-of-the-art graph contrastive learning (GCL) algorithms1 (Table 2).",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
428,
505,
440
],
"spans": [
{
"bbox": [
105,
428,
505,
440
],
"score": 1.0,
"content": "Graph CNE achieved the best results on one of the datasets (PUB), and had close to the best results",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
106,
439,
263,
451
],
"spans": [
{
"bbox": [
106,
439,
263,
451
],
"score": 1.0,
"content": "on other datasets, apart from the ARX.",
"type": "text"
}
],
"index": 14
}
],
"index": 12,
"bbox_fs": [
105,
394,
505,
451
]
},
{
"type": "text",
"bbox": [
107,
455,
505,
576
],
"lines": [
{
"bbox": [
105,
456,
505,
468
],
"spans": [
{
"bbox": [
105,
456,
505,
468
],
"score": 1.0,
"content": "Note that graph CNE was at disadvantage compared to all other GCL methods listed in Table 2",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
467,
505,
479
],
"spans": [
{
"bbox": [
105,
467,
505,
479
],
"score": 1.0,
"content": "because it used an MLP network, whereas other GCL methods traditionally use graph convolutional",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
477,
505,
491
],
"spans": [
{
"bbox": [
105,
477,
505,
491
],
"score": 1.0,
"content": "networks (GCN). GCN takes the entire graph as input and uses message passing, which pulls to-",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
104,
488,
506,
502
],
"spans": [
{
"bbox": [
104,
488,
506,
502
],
"score": 1.0,
"content": "gether embeddings of connected nodes and helps to obtain better embeddings. However, GCN is",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
499,
505,
512
],
"spans": [
{
"bbox": [
105,
499,
505,
512
],
"score": 1.0,
"content": "not able to transform one node at a time, and so a trained GCN cannot be applied to a new, held-out",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
510,
505,
523
],
"spans": [
{
"bbox": [
105,
510,
505,
523
],
"score": 1.0,
"content": "node. In contrast, our graph CNE with MLP can (after training) process one node at a time, which",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
522,
505,
534
],
"spans": [
{
"bbox": [
105,
522,
505,
534
],
"score": 1.0,
"content": "we consider more appropriate for node-level graph learning (see Discussion). There are very few",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
532,
505,
544
],
"spans": [
{
"bbox": [
105,
532,
505,
544
],
"score": 1.0,
"content": "GCL results based on the MLP architecture reported in the literature. Two examples are Local-GCL",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
543,
504,
556
],
"spans": [
{
"bbox": [
105,
543,
504,
556
],
"score": 1.0,
"content": "and GRACE trained with MLP architecture (reported in the OpenReview discussion of Zhang et al.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
554,
505,
568
],
"spans": [
{
"bbox": [
105,
554,
505,
568
],
"score": 1.0,
"content": "(2022) and in Guo et al. (2023) respectively, Table 2): both had lower accuracy compared to our",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
565,
217,
578
],
"spans": [
{
"bbox": [
105,
565,
217,
578
],
"score": 1.0,
"content": "graph CNE on all datasets.",
"type": "text"
}
],
"index": 25
}
],
"index": 20,
"bbox_fs": [
104,
456,
506,
578
]
},
{
"type": "text",
"bbox": [
108,
582,
505,
615
],
"lines": [
{
"bbox": [
105,
581,
505,
595
],
"spans": [
{
"bbox": [
105,
581,
505,
595
],
"score": 1.0,
"content": "For the ARX graph, we did not find any existing MLP-based results. Lower performance of graph",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
593,
505,
606
],
"spans": [
{
"bbox": [
106,
593,
505,
606
],
"score": 1.0,
"content": "CNE compared to GCN-based GCL methods was, again, likely due to the feature space of this graph",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
604,
331,
617
],
"spans": [
{
"bbox": [
105,
604,
331,
617
],
"score": 1.0,
"content": "showing only weak class separation (Table 2, first row).",
"type": "text"
}
],
"index": 28
}
],
"index": 27,
"bbox_fs": [
105,
581,
505,
617
]
},
{
"type": "title",
"bbox": [
108,
635,
190,
648
],
"lines": [
{
"bbox": [
104,
633,
192,
651
],
"spans": [
{
"bbox": [
104,
633,
192,
651
],
"score": 1.0,
"content": "7 DISCUSSION",
"type": "text"
}
],
"index": 29
}
],
"index": 29
},
{
"type": "text",
"bbox": [
107,
662,
443,
673
],
"lines": [
{
"bbox": [
105,
660,
446,
676
],
"spans": [
{
"bbox": [
105,
660,
446,
676
],
"score": 1.0,
"content": "Summary Our paper makes three contributions, two practical and one conceptual:",
"type": "text"
}
],
"index": 30
}
],
"index": 30,
"bbox_fs": [
105,
660,
446,
676
]
},
{
"type": "text",
"bbox": [
114,
685,
504,
707
],
"lines": [
{
"bbox": [
115,
685,
505,
697
],
"spans": [
{
"bbox": [
115,
685,
345,
697
],
"score": 1.0,
"content": "i. We suggested a novel graph layout algorithm, graph",
"type": "text"
},
{
"bbox": [
345,
686,
350,
695
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
350,
685,
505,
697
],
"score": 1.0,
"content": "-SNE, and showed that it outperforms",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
127,
695,
352,
708
],
"spans": [
{
"bbox": [
127,
695,
352,
708
],
"score": 1.0,
"content": "existing competitors in preserving local graph structure.",
"type": "text"
}
],
"index": 32
}
],
"index": 31.5,
"bbox_fs": [
115,
685,
505,
708
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
110,
82,
505,
198
],
"lines": [
{
"bbox": [
113,
82,
505,
96
],
"spans": [
{
"bbox": [
113,
82,
505,
96
],
"score": 1.0,
"content": "ii. We suggested a novel node-level augmentation-free graph contrastive learning algorithm,",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
125,
93,
505,
106
],
"spans": [
{
"bbox": [
125,
93,
505,
106
],
"score": 1.0,
"content": "graph CNE, and showed that it achieves comparable performance to the state-of-the-art meth-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
126,
104,
505,
117
],
"spans": [
{
"bbox": [
126,
104,
505,
117
],
"score": 1.0,
"content": "ods despite using the MLP architecture, and outperforms existing MLP-based graph contrastive",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
126,
115,
194,
128
],
"spans": [
{
"bbox": [
126,
115,
194,
128
],
"score": 1.0,
"content": "learning results.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
110,
141,
505,
155
],
"spans": [
{
"bbox": [
110,
141,
505,
155
],
"score": 1.0,
"content": "iii. We established a conceptual connection between graph layouts and graph contrastive learning:",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
127,
154,
505,
165
],
"spans": [
{
"bbox": [
127,
154,
505,
165
],
"score": 1.0,
"content": "we argued that both are instances of graph embeddings (non-parametric 2D embedding and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
126,
164,
505,
177
],
"spans": [
{
"bbox": [
126,
164,
505,
177
],
"score": 1.0,
"content": "parametric 128D embedding), and both can be efficiently implemented using neighbor em-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
126,
173,
506,
189
],
"spans": [
{
"bbox": [
126,
173,
506,
189
],
"score": 1.0,
"content": "bedding frameworks. We suggested a new task, parametric 2D embeddings (Figure 1), as a",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
128,
185,
320,
199
],
"spans": [
{
"bbox": [
128,
185,
320,
199
],
"score": 1.0,
"content": "‘missing link’ between these two existing tasks.",
"type": "text"
}
],
"index": 8
}
],
"index": 4
},
{
"type": "text",
"bbox": [
107,
223,
505,
300
],
"lines": [
{
"bbox": [
105,
222,
505,
237
],
"spans": [
{
"bbox": [
105,
222,
208,
237
],
"score": 1.0,
"content": "Simplicity Both graph",
"type": "text"
},
{
"bbox": [
208,
224,
213,
233
],
"score": 0.57,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
213,
222,
505,
237
],
"score": 1.0,
"content": "-SNE and graph CNE are remarkably simple, because they use existing",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
234,
505,
246
],
"spans": [
{
"bbox": [
106,
235,
111,
244
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
111,
234,
505,
246
],
"score": 1.0,
"content": "-SNE and CNE machinery out of the box. This is in stark contrast with competing algorithms. For",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
245,
505,
257
],
"spans": [
{
"bbox": [
105,
245,
324,
257
],
"score": 1.0,
"content": "example, existing graph layout algorithms inspired by",
"type": "text"
},
{
"bbox": [
324,
246,
329,
255
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
330,
245,
505,
257
],
"score": 1.0,
"content": "-SNE, such as tsNET (Kruiger et al., 2017),",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
255,
506,
270
],
"spans": [
{
"bbox": [
105,
255,
250,
270
],
"score": 1.0,
"content": "DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
250,
257,
255,
266
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
255,
255,
506,
270
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), all develop their own machinery,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
266,
505,
280
],
"spans": [
{
"bbox": [
105,
266,
335,
280
],
"score": 1.0,
"content": "implementation, and approximations, and deviate from",
"type": "text"
},
{
"bbox": [
335,
268,
340,
277
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
341,
266,
505,
280
],
"score": 1.0,
"content": "-SNE in many different nontrivial ways",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
278,
505,
291
],
"spans": [
{
"bbox": [
105,
278,
348,
291
],
"score": 1.0,
"content": "(see Section 2). However, as we demonstrated, simply using",
"type": "text"
},
{
"bbox": [
349,
279,
353,
288
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
354,
278,
420,
291
],
"score": 1.0,
"content": "-SNE (via graph",
"type": "text"
},
{
"bbox": [
420,
279,
425,
288
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
425,
278,
505,
291
],
"score": 1.0,
"content": "-SNE), outperforms",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
289,
257,
302
],
"spans": [
{
"bbox": [
105,
289,
257,
302
],
"score": 1.0,
"content": "all of them in terms of layout quality.",
"type": "text"
}
],
"index": 15
}
],
"index": 12
},
{
"type": "text",
"bbox": [
107,
305,
505,
405
],
"lines": [
{
"bbox": [
106,
304,
505,
319
],
"spans": [
{
"bbox": [
106,
304,
505,
319
],
"score": 1.0,
"content": "Similarly, in node-level graph contrastive learning (GCL), the focus has been on developing graph",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
317,
506,
330
],
"spans": [
{
"bbox": [
105,
317,
506,
330
],
"score": 1.0,
"content": "augmentations (see Section 2), following the contrastive learning paradigm in computer vision that",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
327,
506,
341
],
"spans": [
{
"bbox": [
105,
327,
506,
341
],
"score": 1.0,
"content": "is based on image augmentations. Augmentation-free GCL methods such as AFGRL (Lee et al.,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
339,
505,
352
],
"spans": [
{
"bbox": [
105,
339,
505,
352
],
"score": 1.0,
"content": "2022) and AF-GCL (Li et al., 2023) instead rely on complex heuristics to select positive pairs. Our",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
350,
506,
362
],
"spans": [
{
"bbox": [
105,
350,
506,
362
],
"score": 1.0,
"content": "approach is conceptually much simpler, as it uses the InfoNCE loss function with graph edges as",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
360,
505,
373
],
"spans": [
{
"bbox": [
105,
360,
505,
373
],
"score": 1.0,
"content": "positive pairs, and nothing else. The closest method in the literature is Local-GCL (Zhang et al.,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
370,
506,
385
],
"spans": [
{
"bbox": [
105,
370,
506,
385
],
"score": 1.0,
"content": "2022), which also uses graph edges as positive pairs. The difference is that Local-GCL uses an",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
381,
506,
396
],
"spans": [
{
"bbox": [
105,
381,
248,
396
],
"score": 1.0,
"content": "approximation scheme to deal with",
"type": "text"
},
{
"bbox": [
248,
381,
275,
394
],
"score": 0.92,
"content": "O ( n ^ { 2 } )",
"type": "inline_equation"
},
{
"bbox": [
276,
381,
506,
396
],
"score": 1.0,
"content": "repulsive forces, whereas we use the standard contrastive",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
394,
381,
407
],
"spans": [
{
"bbox": [
105,
394,
381,
407
],
"score": 1.0,
"content": "learning approach of within-batch repulsion, which is much simpler.",
"type": "text"
}
],
"index": 24
}
],
"index": 20
},
{
"type": "text",
"bbox": [
107,
410,
505,
498
],
"lines": [
{
"bbox": [
105,
410,
505,
423
],
"spans": [
{
"bbox": [
105,
410,
505,
423
],
"score": 1.0,
"content": "All of the existing GCL methods, including Local-GCL, employ graph convolutional neural net-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
421,
506,
434
],
"spans": [
{
"bbox": [
105,
421,
506,
434
],
"score": 1.0,
"content": "works (GCNs). Recent work argued that the reason many GCL algorithms work well has little to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
432,
506,
445
],
"spans": [
{
"bbox": [
105,
432,
506,
445
],
"score": 1.0,
"content": "do with the specific augmentations or heuristics they use, but rather is due to their GCN architecture",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
442,
506,
456
],
"spans": [
{
"bbox": [
105,
442,
506,
456
],
"score": 1.0,
"content": "(Trivedi et al., 2022; Guo et al., 2023). GCN uses message passing between graph nodes, which",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
455,
505,
467
],
"spans": [
{
"bbox": [
106,
455,
505,
467
],
"score": 1.0,
"content": "implicitly makes representations of connected node pairs more similar. In other words, in GCL al-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
465,
506,
477
],
"spans": [
{
"bbox": [
105,
465,
506,
477
],
"score": 1.0,
"content": "gorithms employing GCNs, it is the GCN that does the heavy lifting, and not the specifics of the",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
475,
506,
489
],
"spans": [
{
"bbox": [
105,
475,
506,
489
],
"score": 1.0,
"content": "GCL algorithm. In contrast, our graph CNE uses an MLP network, and nevertheless performed",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
487,
505,
499
],
"spans": [
{
"bbox": [
105,
487,
505,
499
],
"score": 1.0,
"content": "similarly well. See below on why we think MLP is a more suitable choice for node-level GCL tasks.",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5
},
{
"type": "text",
"bbox": [
107,
524,
505,
601
],
"lines": [
{
"bbox": [
105,
524,
505,
537
],
"spans": [
{
"bbox": [
105,
524,
505,
537
],
"score": 1.0,
"content": "Limitations In this work, we focused on complex real-world graphs and have purposefully not",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
534,
505,
548
],
"spans": [
{
"bbox": [
105,
534,
174,
548
],
"score": 1.0,
"content": "tested our graph",
"type": "text"
},
{
"bbox": [
174,
536,
180,
545
],
"score": 0.53,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
180,
534,
505,
548
],
"score": 1.0,
"content": "-SNE on simple planar graphs or 3D mesh graphs that are often used for bench-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
546,
505,
559
],
"spans": [
{
"bbox": [
105,
546,
343,
559
],
"score": 1.0,
"content": "marking graph layout algorithms. We suspect that graph",
"type": "text"
},
{
"bbox": [
343,
547,
349,
556
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
349,
546,
505,
559
],
"score": 1.0,
"content": "-SNE would perform suboptimally on",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
557,
505,
569
],
"spans": [
{
"bbox": [
105,
557,
170,
569
],
"score": 1.0,
"content": "such graphs, as",
"type": "text"
},
{
"bbox": [
170,
558,
175,
567
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
176,
557,
505,
569
],
"score": 1.0,
"content": "-SNE is known to have troubles with embedding simple 2D manifolds such as the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
568,
506,
581
],
"spans": [
{
"bbox": [
105,
568,
506,
581
],
"score": 1.0,
"content": "Swiss roll. To some extent this can be addressed by increasing the degree of freedom parameter of",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
579,
505,
591
],
"spans": [
{
"bbox": [
105,
579,
121,
591
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
121,
580,
126,
589
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
126,
579,
505,
591
],
"score": 1.0,
"content": "-distribution or using the Gaussian kernel instead (Kobak et al., 2019), and/or by increasing the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
590,
450,
602
],
"spans": [
{
"bbox": [
105,
590,
450,
602
],
"score": 1.0,
"content": "exaggeration value (Kobak & Berens, 2019; Bohm et al., 2022; Damrich et al., 2022). ¨",
"type": "text"
}
],
"index": 39
}
],
"index": 36
},
{
"type": "text",
"bbox": [
107,
606,
504,
673
],
"lines": [
{
"bbox": [
105,
606,
506,
619
],
"spans": [
{
"bbox": [
105,
606,
506,
619
],
"score": 1.0,
"content": "Our graph CNE relies on the MLP and we did not experiment with GCN architecture. This, however,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
618,
506,
631
],
"spans": [
{
"bbox": [
105,
618,
506,
631
],
"score": 1.0,
"content": "is not a limitation but a purposeful design choice: we think that GCN, whereas very meaningful",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
628,
505,
641
],
"spans": [
{
"bbox": [
105,
628,
505,
641
],
"score": 1.0,
"content": "for graph-level learning, is less applicable for node-level learning, where one may want to apply",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
640,
506,
651
],
"spans": [
{
"bbox": [
106,
640,
506,
651
],
"score": 1.0,
"content": "the trained model to a set of new objects (based on their node features). With GCN, this is not",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
650,
506,
663
],
"spans": [
{
"bbox": [
105,
650,
506,
663
],
"score": 1.0,
"content": "possible, as it requires the entire graph to be passed in at the same time. We therefore consider MLP",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
662,
310,
673
],
"spans": [
{
"bbox": [
105,
662,
310,
673
],
"score": 1.0,
"content": "architecture more appropriate for node-level GCL.",
"type": "text"
}
],
"index": 45
}
],
"index": 42.5
},
{
"type": "text",
"bbox": [
108,
699,
504,
731
],
"lines": [
{
"bbox": [
105,
698,
505,
712
],
"spans": [
{
"bbox": [
105,
698,
505,
712
],
"score": 1.0,
"content": "Take-home message We showed that graph layouts and graph contrastive learning are intimately",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
709,
505,
721
],
"spans": [
{
"bbox": [
106,
709,
505,
721
],
"score": 1.0,
"content": "related and can be approached by existing neighbour embedding frameworks, surpassing state-of-",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
720,
166,
732
],
"spans": [
{
"bbox": [
106,
720,
166,
732
],
"score": 1.0,
"content": "the-art results.",
"type": "text"
}
],
"index": 48
}
],
"index": 47
}
],
"page_idx": 8,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
302,
751,
308,
759
],
"lines": [
{
"bbox": [
302,
751,
309,
762
],
"spans": [
{
"bbox": [
302,
751,
309,
762
],
"score": 1.0,
"content": "9",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
110,
82,
505,
198
],
"lines": [
{
"bbox": [
113,
82,
505,
96
],
"spans": [
{
"bbox": [
113,
82,
505,
96
],
"score": 1.0,
"content": "ii. We suggested a novel node-level augmentation-free graph contrastive learning algorithm,",
"type": "text"
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
125,
93,
505,
106
],
"spans": [
{
"bbox": [
125,
93,
505,
106
],
"score": 1.0,
"content": "graph CNE, and showed that it achieves comparable performance to the state-of-the-art meth-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
126,
104,
505,
117
],
"spans": [
{
"bbox": [
126,
104,
505,
117
],
"score": 1.0,
"content": "ods despite using the MLP architecture, and outperforms existing MLP-based graph contrastive",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
126,
115,
194,
128
],
"spans": [
{
"bbox": [
126,
115,
194,
128
],
"score": 1.0,
"content": "learning results.",
"type": "text"
}
],
"index": 3,
"is_list_end_line": true
},
{
"bbox": [
110,
141,
505,
155
],
"spans": [
{
"bbox": [
110,
141,
505,
155
],
"score": 1.0,
"content": "iii. We established a conceptual connection between graph layouts and graph contrastive learning:",
"type": "text"
}
],
"index": 4,
"is_list_start_line": true
},
{
"bbox": [
127,
154,
505,
165
],
"spans": [
{
"bbox": [
127,
154,
505,
165
],
"score": 1.0,
"content": "we argued that both are instances of graph embeddings (non-parametric 2D embedding and",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
126,
164,
505,
177
],
"spans": [
{
"bbox": [
126,
164,
505,
177
],
"score": 1.0,
"content": "parametric 128D embedding), and both can be efficiently implemented using neighbor em-",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
126,
173,
506,
189
],
"spans": [
{
"bbox": [
126,
173,
506,
189
],
"score": 1.0,
"content": "bedding frameworks. We suggested a new task, parametric 2D embeddings (Figure 1), as a",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
128,
185,
320,
199
],
"spans": [
{
"bbox": [
128,
185,
320,
199
],
"score": 1.0,
"content": "‘missing link’ between these two existing tasks.",
"type": "text"
}
],
"index": 8,
"is_list_end_line": true
}
],
"index": 4,
"bbox_fs": [
110,
82,
506,
199
]
},
{
"type": "text",
"bbox": [
107,
223,
505,
300
],
"lines": [
{
"bbox": [
105,
222,
505,
237
],
"spans": [
{
"bbox": [
105,
222,
208,
237
],
"score": 1.0,
"content": "Simplicity Both graph",
"type": "text"
},
{
"bbox": [
208,
224,
213,
233
],
"score": 0.57,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
213,
222,
505,
237
],
"score": 1.0,
"content": "-SNE and graph CNE are remarkably simple, because they use existing",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
106,
234,
505,
246
],
"spans": [
{
"bbox": [
106,
235,
111,
244
],
"score": 0.7,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
111,
234,
505,
246
],
"score": 1.0,
"content": "-SNE and CNE machinery out of the box. This is in stark contrast with competing algorithms. For",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
245,
505,
257
],
"spans": [
{
"bbox": [
105,
245,
324,
257
],
"score": 1.0,
"content": "example, existing graph layout algorithms inspired by",
"type": "text"
},
{
"bbox": [
324,
246,
329,
255
],
"score": 0.68,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
330,
245,
505,
257
],
"score": 1.0,
"content": "-SNE, such as tsNET (Kruiger et al., 2017),",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
255,
506,
270
],
"spans": [
{
"bbox": [
105,
255,
250,
270
],
"score": 1.0,
"content": "DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
250,
257,
255,
266
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
255,
255,
506,
270
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), all develop their own machinery,",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
266,
505,
280
],
"spans": [
{
"bbox": [
105,
266,
335,
280
],
"score": 1.0,
"content": "implementation, and approximations, and deviate from",
"type": "text"
},
{
"bbox": [
335,
268,
340,
277
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
341,
266,
505,
280
],
"score": 1.0,
"content": "-SNE in many different nontrivial ways",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
278,
505,
291
],
"spans": [
{
"bbox": [
105,
278,
348,
291
],
"score": 1.0,
"content": "(see Section 2). However, as we demonstrated, simply using",
"type": "text"
},
{
"bbox": [
349,
279,
353,
288
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
354,
278,
420,
291
],
"score": 1.0,
"content": "-SNE (via graph",
"type": "text"
},
{
"bbox": [
420,
279,
425,
288
],
"score": 0.58,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
425,
278,
505,
291
],
"score": 1.0,
"content": "-SNE), outperforms",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
289,
257,
302
],
"spans": [
{
"bbox": [
105,
289,
257,
302
],
"score": 1.0,
"content": "all of them in terms of layout quality.",
"type": "text"
}
],
"index": 15
}
],
"index": 12,
"bbox_fs": [
105,
222,
506,
302
]
},
{
"type": "text",
"bbox": [
107,
305,
505,
405
],
"lines": [
{
"bbox": [
106,
304,
505,
319
],
"spans": [
{
"bbox": [
106,
304,
505,
319
],
"score": 1.0,
"content": "Similarly, in node-level graph contrastive learning (GCL), the focus has been on developing graph",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
317,
506,
330
],
"spans": [
{
"bbox": [
105,
317,
506,
330
],
"score": 1.0,
"content": "augmentations (see Section 2), following the contrastive learning paradigm in computer vision that",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
327,
506,
341
],
"spans": [
{
"bbox": [
105,
327,
506,
341
],
"score": 1.0,
"content": "is based on image augmentations. Augmentation-free GCL methods such as AFGRL (Lee et al.,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
105,
339,
505,
352
],
"spans": [
{
"bbox": [
105,
339,
505,
352
],
"score": 1.0,
"content": "2022) and AF-GCL (Li et al., 2023) instead rely on complex heuristics to select positive pairs. Our",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
350,
506,
362
],
"spans": [
{
"bbox": [
105,
350,
506,
362
],
"score": 1.0,
"content": "approach is conceptually much simpler, as it uses the InfoNCE loss function with graph edges as",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
360,
505,
373
],
"spans": [
{
"bbox": [
105,
360,
505,
373
],
"score": 1.0,
"content": "positive pairs, and nothing else. The closest method in the literature is Local-GCL (Zhang et al.,",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
105,
370,
506,
385
],
"spans": [
{
"bbox": [
105,
370,
506,
385
],
"score": 1.0,
"content": "2022), which also uses graph edges as positive pairs. The difference is that Local-GCL uses an",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
381,
506,
396
],
"spans": [
{
"bbox": [
105,
381,
248,
396
],
"score": 1.0,
"content": "approximation scheme to deal with",
"type": "text"
},
{
"bbox": [
248,
381,
275,
394
],
"score": 0.92,
"content": "O ( n ^ { 2 } )",
"type": "inline_equation"
},
{
"bbox": [
276,
381,
506,
396
],
"score": 1.0,
"content": "repulsive forces, whereas we use the standard contrastive",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
394,
381,
407
],
"spans": [
{
"bbox": [
105,
394,
381,
407
],
"score": 1.0,
"content": "learning approach of within-batch repulsion, which is much simpler.",
"type": "text"
}
],
"index": 24
}
],
"index": 20,
"bbox_fs": [
105,
304,
506,
407
]
},
{
"type": "text",
"bbox": [
107,
410,
505,
498
],
"lines": [
{
"bbox": [
105,
410,
505,
423
],
"spans": [
{
"bbox": [
105,
410,
505,
423
],
"score": 1.0,
"content": "All of the existing GCL methods, including Local-GCL, employ graph convolutional neural net-",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
421,
506,
434
],
"spans": [
{
"bbox": [
105,
421,
506,
434
],
"score": 1.0,
"content": "works (GCNs). Recent work argued that the reason many GCL algorithms work well has little to",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
105,
432,
506,
445
],
"spans": [
{
"bbox": [
105,
432,
506,
445
],
"score": 1.0,
"content": "do with the specific augmentations or heuristics they use, but rather is due to their GCN architecture",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
442,
506,
456
],
"spans": [
{
"bbox": [
105,
442,
506,
456
],
"score": 1.0,
"content": "(Trivedi et al., 2022; Guo et al., 2023). GCN uses message passing between graph nodes, which",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
106,
455,
505,
467
],
"spans": [
{
"bbox": [
106,
455,
505,
467
],
"score": 1.0,
"content": "implicitly makes representations of connected node pairs more similar. In other words, in GCL al-",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
105,
465,
506,
477
],
"spans": [
{
"bbox": [
105,
465,
506,
477
],
"score": 1.0,
"content": "gorithms employing GCNs, it is the GCN that does the heavy lifting, and not the specifics of the",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
475,
506,
489
],
"spans": [
{
"bbox": [
105,
475,
506,
489
],
"score": 1.0,
"content": "GCL algorithm. In contrast, our graph CNE uses an MLP network, and nevertheless performed",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
105,
487,
505,
499
],
"spans": [
{
"bbox": [
105,
487,
505,
499
],
"score": 1.0,
"content": "similarly well. See below on why we think MLP is a more suitable choice for node-level GCL tasks.",
"type": "text"
}
],
"index": 32
}
],
"index": 28.5,
"bbox_fs": [
105,
410,
506,
499
]
},
{
"type": "text",
"bbox": [
107,
524,
505,
601
],
"lines": [
{
"bbox": [
105,
524,
505,
537
],
"spans": [
{
"bbox": [
105,
524,
505,
537
],
"score": 1.0,
"content": "Limitations In this work, we focused on complex real-world graphs and have purposefully not",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
105,
534,
505,
548
],
"spans": [
{
"bbox": [
105,
534,
174,
548
],
"score": 1.0,
"content": "tested our graph",
"type": "text"
},
{
"bbox": [
174,
536,
180,
545
],
"score": 0.53,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
180,
534,
505,
548
],
"score": 1.0,
"content": "-SNE on simple planar graphs or 3D mesh graphs that are often used for bench-",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
546,
505,
559
],
"spans": [
{
"bbox": [
105,
546,
343,
559
],
"score": 1.0,
"content": "marking graph layout algorithms. We suspect that graph",
"type": "text"
},
{
"bbox": [
343,
547,
349,
556
],
"score": 0.65,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
349,
546,
505,
559
],
"score": 1.0,
"content": "-SNE would perform suboptimally on",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
105,
557,
505,
569
],
"spans": [
{
"bbox": [
105,
557,
170,
569
],
"score": 1.0,
"content": "such graphs, as",
"type": "text"
},
{
"bbox": [
170,
558,
175,
567
],
"score": 0.69,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
176,
557,
505,
569
],
"score": 1.0,
"content": "-SNE is known to have troubles with embedding simple 2D manifolds such as the",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
568,
506,
581
],
"spans": [
{
"bbox": [
105,
568,
506,
581
],
"score": 1.0,
"content": "Swiss roll. To some extent this can be addressed by increasing the degree of freedom parameter of",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
105,
579,
505,
591
],
"spans": [
{
"bbox": [
105,
579,
121,
591
],
"score": 1.0,
"content": "the",
"type": "text"
},
{
"bbox": [
121,
580,
126,
589
],
"score": 0.74,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
126,
579,
505,
591
],
"score": 1.0,
"content": "-distribution or using the Gaussian kernel instead (Kobak et al., 2019), and/or by increasing the",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
105,
590,
450,
602
],
"spans": [
{
"bbox": [
105,
590,
450,
602
],
"score": 1.0,
"content": "exaggeration value (Kobak & Berens, 2019; Bohm et al., 2022; Damrich et al., 2022). ¨",
"type": "text"
}
],
"index": 39
}
],
"index": 36,
"bbox_fs": [
105,
524,
506,
602
]
},
{
"type": "text",
"bbox": [
107,
606,
504,
673
],
"lines": [
{
"bbox": [
105,
606,
506,
619
],
"spans": [
{
"bbox": [
105,
606,
506,
619
],
"score": 1.0,
"content": "Our graph CNE relies on the MLP and we did not experiment with GCN architecture. This, however,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
105,
618,
506,
631
],
"spans": [
{
"bbox": [
105,
618,
506,
631
],
"score": 1.0,
"content": "is not a limitation but a purposeful design choice: we think that GCN, whereas very meaningful",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
105,
628,
505,
641
],
"spans": [
{
"bbox": [
105,
628,
505,
641
],
"score": 1.0,
"content": "for graph-level learning, is less applicable for node-level learning, where one may want to apply",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
106,
640,
506,
651
],
"spans": [
{
"bbox": [
106,
640,
506,
651
],
"score": 1.0,
"content": "the trained model to a set of new objects (based on their node features). With GCN, this is not",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
105,
650,
506,
663
],
"spans": [
{
"bbox": [
105,
650,
506,
663
],
"score": 1.0,
"content": "possible, as it requires the entire graph to be passed in at the same time. We therefore consider MLP",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
105,
662,
310,
673
],
"spans": [
{
"bbox": [
105,
662,
310,
673
],
"score": 1.0,
"content": "architecture more appropriate for node-level GCL.",
"type": "text"
}
],
"index": 45
}
],
"index": 42.5,
"bbox_fs": [
105,
606,
506,
673
]
},
{
"type": "text",
"bbox": [
108,
699,
504,
731
],
"lines": [
{
"bbox": [
105,
698,
505,
712
],
"spans": [
{
"bbox": [
105,
698,
505,
712
],
"score": 1.0,
"content": "Take-home message We showed that graph layouts and graph contrastive learning are intimately",
"type": "text"
}
],
"index": 46
},
{
"bbox": [
106,
709,
505,
721
],
"spans": [
{
"bbox": [
106,
709,
505,
721
],
"score": 1.0,
"content": "related and can be approached by existing neighbour embedding frameworks, surpassing state-of-",
"type": "text"
}
],
"index": 47
},
{
"bbox": [
106,
720,
166,
732
],
"spans": [
{
"bbox": [
106,
720,
166,
732
],
"score": 1.0,
"content": "the-art results.",
"type": "text"
}
],
"index": 48
}
],
"index": 47,
"bbox_fs": [
105,
698,
505,
732
]
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
81,
175,
93
],
"lines": [
{
"bbox": [
106,
81,
176,
95
],
"spans": [
{
"bbox": [
106,
81,
176,
95
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
100,
503,
123
],
"lines": [
{
"bbox": [
106,
99,
505,
112
],
"spans": [
{
"bbox": [
106,
99,
505,
112
],
"score": 1.0,
"content": "Aleksandr Artemenkov and Maxim Panov. NCVis: noise contrastive approach for scalable visual-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
116,
110,
421,
123
],
"spans": [
{
"bbox": [
116,
110,
421,
123
],
"score": 1.0,
"content": "ization. In Proceedings of The Web Conference 2020, pp. 2941–2947, 2020.",
"type": "text"
}
],
"index": 2
}
],
"index": 1.5
},
{
"type": "text",
"bbox": [
100,
129,
506,
754
],
"lines": [
{
"bbox": [
105,
129,
506,
142
],
"spans": [
{
"bbox": [
105,
129,
506,
142
],
"score": 1.0,
"content": "Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
114,
140,
365,
154
],
"spans": [
{
"bbox": [
114,
140,
365,
154
],
"score": 1.0,
"content": "representation. Neural Computation, 15(6):1373–1396, 2003.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
105,
158,
504,
171
],
"spans": [
{
"bbox": [
105,
158,
504,
171
],
"score": 1.0,
"content": "Anna C Belkina, Christopher O Ciccolella, Rina Anno, Richard Halpert, Josef Spidlen, and Jen-",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
114,
168,
506,
183
],
"spans": [
{
"bbox": [
114,
168,
506,
183
],
"score": 1.0,
"content": "nifer E Snyder-Cappione. Automated optimized parameters for T-distributed stochastic neighbor",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
116,
181,
505,
194
],
"spans": [
{
"bbox": [
116,
181,
505,
194
],
"score": 1.0,
"content": "embedding improve visualization and analysis of large datasets. Nature Communications, 10(1):",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
115,
191,
168,
204
],
"spans": [
{
"bbox": [
115,
191,
168,
204
],
"score": 1.0,
"content": "5415, 2019.",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
211,
505,
224
],
"spans": [
{
"bbox": [
105,
211,
505,
224
],
"score": 1.0,
"content": "Jan Niklas Bohm, Philipp Berens, and Dmitry Kobak. Attraction-repulsion spectrum in neighbor ¨",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
115,
222,
447,
235
],
"spans": [
{
"bbox": [
115,
222,
447,
235
],
"score": 1.0,
"content": "embeddings. The Journal of Machine Learning Research, 23(1):4118–4149, 2022.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
240,
506,
253
],
"spans": [
{
"bbox": [
105,
240,
506,
253
],
"score": 1.0,
"content": "Jan Niklas Bohm, Philipp Berens, and Dmitry Kobak. Unsupervised visualization of image datasets ¨",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
116,
252,
473,
264
],
"spans": [
{
"bbox": [
116,
252,
473,
264
],
"score": 1.0,
"content": "using contrastive learning. International Conference on Learning Representations, 2023.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
105,
268,
506,
282
],
"spans": [
{
"bbox": [
105,
268,
506,
282
],
"score": 1.0,
"content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
114,
279,
506,
295
],
"spans": [
{
"bbox": [
114,
279,
506,
295
],
"score": 1.0,
"content": "contrastive learning of visual representations. In International Conference on Machine Learning,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
114,
291,
239,
304
],
"spans": [
{
"bbox": [
114,
291,
239,
304
],
"score": 1.0,
"content": "pp. 1597–1607. PMLR, 2020.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
309,
505,
322
],
"spans": [
{
"bbox": [
105,
309,
434,
322
],
"score": 1.0,
"content": "Sebastian Damrich, Niklas Bohm, Fred A Hamprecht, and Dmitry Kobak. From ¨",
"type": "text"
},
{
"bbox": [
434,
311,
439,
319
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
440,
309,
505,
322
],
"score": 1.0,
"content": "-SNE to UMAP",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
115,
321,
506,
334
],
"spans": [
{
"bbox": [
115,
321,
506,
334
],
"score": 1.0,
"content": "with contrastive learning. In The Eleventh International Conference on Learning Representations,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
115,
331,
142,
345
],
"spans": [
{
"bbox": [
115,
331,
142,
345
],
"score": 1.0,
"content": "2022.",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
104,
348,
505,
364
],
"spans": [
{
"bbox": [
104,
348,
505,
364
],
"score": 1.0,
"content": "Thomas MJ Fruchterman and Edward M Reingold. Graph drawing by force-directed placement.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
116,
361,
366,
374
],
"spans": [
{
"bbox": [
116,
361,
366,
374
],
"score": 1.0,
"content": "Software: Practice and Experience, 21(11):1129–1164, 1991.",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
105,
379,
506,
392
],
"spans": [
{
"bbox": [
105,
379,
506,
392
],
"score": 1.0,
"content": "Xiaojun Guo, Yifei Wang, Zeming Wei, and Yisen Wang. Architecture matters: Uncovering implicit",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
115,
390,
505,
404
],
"spans": [
{
"bbox": [
115,
390,
505,
404
],
"score": 1.0,
"content": "mechanisms in graph contrastive learning. In Thirty-seventh Conference on Neural Information",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
115,
402,
225,
415
],
"spans": [
{
"bbox": [
115,
402,
225,
415
],
"score": 1.0,
"content": "Processing Systems, 2023.",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
105,
420,
505,
433
],
"spans": [
{
"bbox": [
105,
420,
505,
433
],
"score": 1.0,
"content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
114,
430,
506,
444
],
"spans": [
{
"bbox": [
114,
430,
506,
444
],
"score": 1.0,
"content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
113,
439,
506,
457
],
"spans": [
{
"bbox": [
113,
439,
506,
457
],
"score": 1.0,
"content": "Artificial Intelligence and Statistics, pp. 297–304. JMLR Workshop and Conference Proceedings,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
451,
142,
466
],
"spans": [
{
"bbox": [
115,
451,
142,
466
],
"score": 1.0,
"content": "2010.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
104,
469,
505,
485
],
"spans": [
{
"bbox": [
104,
469,
505,
485
],
"score": 1.0,
"content": "Aric Hagberg, Pieter Swart, and Daniel S Chult. Exploring network structure, dynamics, and func-",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
482,
409,
495
],
"spans": [
{
"bbox": [
115,
482,
409,
495
],
"score": 1.0,
"content": "tion using NetworkX. Technical report, Los Alamos National Lab, 2008.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
104,
499,
506,
515
],
"spans": [
{
"bbox": [
104,
499,
506,
515
],
"score": 1.0,
"content": "Kaveh Hassani and Amir Hosein Khasahmadi. Contrastive multi-view representation learning on",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
114,
511,
477,
525
],
"spans": [
{
"bbox": [
114,
511,
477,
525
],
"score": 1.0,
"content": "graphs. In International Conference on Machine Learning, pp. 4116–4126. PMLR, 2020.",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
104,
528,
505,
544
],
"spans": [
{
"bbox": [
104,
528,
505,
544
],
"score": 1.0,
"content": "Geoffrey E Hinton and Sam Roweis. Stochastic neighbor embedding. Advances in Neural Informa-",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
115,
541,
257,
554
],
"spans": [
{
"bbox": [
115,
541,
257,
554
],
"score": 1.0,
"content": "tion Processing Systems, 15, 2002.",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
104,
557,
506,
574
],
"spans": [
{
"bbox": [
104,
557,
506,
574
],
"score": 1.0,
"content": "Tianyang Hu, Zhili Liu, Fengwei Zhou, Wenjia Wang, and Weiran Huang. Your contrastive learning",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
114,
569,
505,
585
],
"spans": [
{
"bbox": [
114,
569,
505,
585
],
"score": 1.0,
"content": "is secretly doing stochastic neighbor embedding. International Conference on Learning Repre-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
114,
582,
187,
593
],
"spans": [
{
"bbox": [
114,
582,
187,
593
],
"score": 1.0,
"content": "sentations, 2023.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
598,
506,
614
],
"spans": [
{
"bbox": [
105,
598,
506,
614
],
"score": 1.0,
"content": "Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta,",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
114,
609,
506,
624
],
"spans": [
{
"bbox": [
114,
609,
506,
624
],
"score": 1.0,
"content": "and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. Advances",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
623,
386,
634
],
"spans": [
{
"bbox": [
116,
623,
386,
634
],
"score": 1.0,
"content": "in Neural Information Processing Systems, 33:22118–22133, 2020.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
638,
506,
654
],
"spans": [
{
"bbox": [
105,
638,
506,
654
],
"score": 1.0,
"content": "Yifan Hu. Efficient, high-quality force-directed graph drawing. Mathematica Journal, 10(1):37–71,",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
114,
650,
143,
665
],
"spans": [
{
"bbox": [
114,
650,
143,
665
],
"score": 1.0,
"content": "2005.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
103,
668,
507,
684
],
"spans": [
{
"bbox": [
103,
668,
507,
684
],
"score": 1.0,
"content": "Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. ForceAtlas2, a",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
115,
681,
505,
694
],
"spans": [
{
"bbox": [
115,
681,
505,
694
],
"score": 1.0,
"content": "continuous graph layout algorithm for handy network visualization designed for the gephi soft-",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
692,
264,
704
],
"spans": [
{
"bbox": [
115,
692,
264,
704
],
"score": 1.0,
"content": "ware. PloS One, 9(6):e98679, 2014.",
"type": "text"
}
],
"index": 44
},
{
"bbox": [
104,
708,
506,
724
],
"spans": [
{
"bbox": [
104,
708,
506,
724
],
"score": 1.0,
"content": "Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the",
"type": "text"
}
],
"index": 45
},
{
"bbox": [
115,
720,
398,
735
],
"spans": [
{
"bbox": [
115,
720,
398,
735
],
"score": 1.0,
"content": "limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.",
"type": "text"
}
],
"index": 46
}
],
"index": 24.5
}
],
"page_idx": 9,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
301,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
313,
765
],
"spans": [
{
"bbox": [
299,
750,
313,
765
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 14
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
81,
175,
93
],
"lines": [
{
"bbox": [
106,
81,
176,
95
],
"spans": [
{
"bbox": [
106,
81,
176,
95
],
"score": 1.0,
"content": "REFERENCES",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "text",
"bbox": [
107,
100,
503,
123
],
"lines": [
{
"bbox": [
106,
99,
505,
112
],
"spans": [
{
"bbox": [
106,
99,
505,
112
],
"score": 1.0,
"content": "Aleksandr Artemenkov and Maxim Panov. NCVis: noise contrastive approach for scalable visual-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
116,
110,
421,
123
],
"spans": [
{
"bbox": [
116,
110,
421,
123
],
"score": 1.0,
"content": "ization. In Proceedings of The Web Conference 2020, pp. 2941–2947, 2020.",
"type": "text"
}
],
"index": 2
}
],
"index": 1.5,
"bbox_fs": [
106,
99,
505,
123
]
},
{
"type": "list",
"bbox": [
100,
129,
506,
754
],
"lines": [
{
"bbox": [
105,
129,
506,
142
],
"spans": [
{
"bbox": [
105,
129,
506,
142
],
"score": 1.0,
"content": "Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps for dimensionality reduction and data",
"type": "text"
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
114,
140,
365,
154
],
"spans": [
{
"bbox": [
114,
140,
365,
154
],
"score": 1.0,
"content": "representation. Neural Computation, 15(6):1373–1396, 2003.",
"type": "text"
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
105,
158,
504,
171
],
"spans": [
{
"bbox": [
105,
158,
504,
171
],
"score": 1.0,
"content": "Anna C Belkina, Christopher O Ciccolella, Rina Anno, Richard Halpert, Josef Spidlen, and Jen-",
"type": "text"
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
114,
168,
506,
183
],
"spans": [
{
"bbox": [
114,
168,
506,
183
],
"score": 1.0,
"content": "nifer E Snyder-Cappione. Automated optimized parameters for T-distributed stochastic neighbor",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
116,
181,
505,
194
],
"spans": [
{
"bbox": [
116,
181,
505,
194
],
"score": 1.0,
"content": "embedding improve visualization and analysis of large datasets. Nature Communications, 10(1):",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
115,
191,
168,
204
],
"spans": [
{
"bbox": [
115,
191,
168,
204
],
"score": 1.0,
"content": "5415, 2019.",
"type": "text"
}
],
"index": 8,
"is_list_end_line": true
},
{
"bbox": [
105,
211,
505,
224
],
"spans": [
{
"bbox": [
105,
211,
505,
224
],
"score": 1.0,
"content": "Jan Niklas Bohm, Philipp Berens, and Dmitry Kobak. Attraction-repulsion spectrum in neighbor ¨",
"type": "text"
}
],
"index": 9,
"is_list_start_line": true
},
{
"bbox": [
115,
222,
447,
235
],
"spans": [
{
"bbox": [
115,
222,
447,
235
],
"score": 1.0,
"content": "embeddings. The Journal of Machine Learning Research, 23(1):4118–4149, 2022.",
"type": "text"
}
],
"index": 10,
"is_list_end_line": true
},
{
"bbox": [
105,
240,
506,
253
],
"spans": [
{
"bbox": [
105,
240,
506,
253
],
"score": 1.0,
"content": "Jan Niklas Bohm, Philipp Berens, and Dmitry Kobak. Unsupervised visualization of image datasets ¨",
"type": "text"
}
],
"index": 11,
"is_list_start_line": true
},
{
"bbox": [
116,
252,
473,
264
],
"spans": [
{
"bbox": [
116,
252,
473,
264
],
"score": 1.0,
"content": "using contrastive learning. International Conference on Learning Representations, 2023.",
"type": "text"
}
],
"index": 12,
"is_list_end_line": true
},
{
"bbox": [
105,
268,
506,
282
],
"spans": [
{
"bbox": [
105,
268,
506,
282
],
"score": 1.0,
"content": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for",
"type": "text"
}
],
"index": 13,
"is_list_start_line": true
},
{
"bbox": [
114,
279,
506,
295
],
"spans": [
{
"bbox": [
114,
279,
506,
295
],
"score": 1.0,
"content": "contrastive learning of visual representations. In International Conference on Machine Learning,",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
114,
291,
239,
304
],
"spans": [
{
"bbox": [
114,
291,
239,
304
],
"score": 1.0,
"content": "pp. 1597–1607. PMLR, 2020.",
"type": "text"
}
],
"index": 15,
"is_list_end_line": true
},
{
"bbox": [
105,
309,
505,
322
],
"spans": [
{
"bbox": [
105,
309,
434,
322
],
"score": 1.0,
"content": "Sebastian Damrich, Niklas Bohm, Fred A Hamprecht, and Dmitry Kobak. From ¨",
"type": "text"
},
{
"bbox": [
434,
311,
439,
319
],
"score": 0.6,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
440,
309,
505,
322
],
"score": 1.0,
"content": "-SNE to UMAP",
"type": "text"
}
],
"index": 16,
"is_list_start_line": true
},
{
"bbox": [
115,
321,
506,
334
],
"spans": [
{
"bbox": [
115,
321,
506,
334
],
"score": 1.0,
"content": "with contrastive learning. In The Eleventh International Conference on Learning Representations,",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
115,
331,
142,
345
],
"spans": [
{
"bbox": [
115,
331,
142,
345
],
"score": 1.0,
"content": "2022.",
"type": "text"
}
],
"index": 18,
"is_list_end_line": true
},
{
"bbox": [
104,
348,
505,
364
],
"spans": [
{
"bbox": [
104,
348,
505,
364
],
"score": 1.0,
"content": "Thomas MJ Fruchterman and Edward M Reingold. Graph drawing by force-directed placement.",
"type": "text"
}
],
"index": 19,
"is_list_start_line": true
},
{
"bbox": [
116,
361,
366,
374
],
"spans": [
{
"bbox": [
116,
361,
366,
374
],
"score": 1.0,
"content": "Software: Practice and Experience, 21(11):1129–1164, 1991.",
"type": "text"
}
],
"index": 20,
"is_list_end_line": true
},
{
"bbox": [
105,
379,
506,
392
],
"spans": [
{
"bbox": [
105,
379,
506,
392
],
"score": 1.0,
"content": "Xiaojun Guo, Yifei Wang, Zeming Wei, and Yisen Wang. Architecture matters: Uncovering implicit",
"type": "text"
}
],
"index": 21,
"is_list_start_line": true
},
{
"bbox": [
115,
390,
505,
404
],
"spans": [
{
"bbox": [
115,
390,
505,
404
],
"score": 1.0,
"content": "mechanisms in graph contrastive learning. In Thirty-seventh Conference on Neural Information",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
115,
402,
225,
415
],
"spans": [
{
"bbox": [
115,
402,
225,
415
],
"score": 1.0,
"content": "Processing Systems, 2023.",
"type": "text"
}
],
"index": 23,
"is_list_end_line": true
},
{
"bbox": [
105,
420,
505,
433
],
"spans": [
{
"bbox": [
105,
420,
505,
433
],
"score": 1.0,
"content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨",
"type": "text"
}
],
"index": 24,
"is_list_start_line": true
},
{
"bbox": [
114,
430,
506,
444
],
"spans": [
{
"bbox": [
114,
430,
506,
444
],
"score": 1.0,
"content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
113,
439,
506,
457
],
"spans": [
{
"bbox": [
113,
439,
506,
457
],
"score": 1.0,
"content": "Artificial Intelligence and Statistics, pp. 297–304. JMLR Workshop and Conference Proceedings,",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
451,
142,
466
],
"spans": [
{
"bbox": [
115,
451,
142,
466
],
"score": 1.0,
"content": "2010.",
"type": "text"
}
],
"index": 27,
"is_list_end_line": true
},
{
"bbox": [
104,
469,
505,
485
],
"spans": [
{
"bbox": [
104,
469,
505,
485
],
"score": 1.0,
"content": "Aric Hagberg, Pieter Swart, and Daniel S Chult. Exploring network structure, dynamics, and func-",
"type": "text"
}
],
"index": 28,
"is_list_start_line": true
},
{
"bbox": [
115,
482,
409,
495
],
"spans": [
{
"bbox": [
115,
482,
409,
495
],
"score": 1.0,
"content": "tion using NetworkX. Technical report, Los Alamos National Lab, 2008.",
"type": "text"
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
104,
499,
506,
515
],
"spans": [
{
"bbox": [
104,
499,
506,
515
],
"score": 1.0,
"content": "Kaveh Hassani and Amir Hosein Khasahmadi. Contrastive multi-view representation learning on",
"type": "text"
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
114,
511,
477,
525
],
"spans": [
{
"bbox": [
114,
511,
477,
525
],
"score": 1.0,
"content": "graphs. In International Conference on Machine Learning, pp. 4116–4126. PMLR, 2020.",
"type": "text"
}
],
"index": 31,
"is_list_end_line": true
},
{
"bbox": [
104,
528,
505,
544
],
"spans": [
{
"bbox": [
104,
528,
505,
544
],
"score": 1.0,
"content": "Geoffrey E Hinton and Sam Roweis. Stochastic neighbor embedding. Advances in Neural Informa-",
"type": "text"
}
],
"index": 32,
"is_list_start_line": true
},
{
"bbox": [
115,
541,
257,
554
],
"spans": [
{
"bbox": [
115,
541,
257,
554
],
"score": 1.0,
"content": "tion Processing Systems, 15, 2002.",
"type": "text"
}
],
"index": 33,
"is_list_end_line": true
},
{
"bbox": [
104,
557,
506,
574
],
"spans": [
{
"bbox": [
104,
557,
506,
574
],
"score": 1.0,
"content": "Tianyang Hu, Zhili Liu, Fengwei Zhou, Wenjia Wang, and Weiran Huang. Your contrastive learning",
"type": "text"
}
],
"index": 34,
"is_list_start_line": true
},
{
"bbox": [
114,
569,
505,
585
],
"spans": [
{
"bbox": [
114,
569,
505,
585
],
"score": 1.0,
"content": "is secretly doing stochastic neighbor embedding. International Conference on Learning Repre-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
114,
582,
187,
593
],
"spans": [
{
"bbox": [
114,
582,
187,
593
],
"score": 1.0,
"content": "sentations, 2023.",
"type": "text"
}
],
"index": 36,
"is_list_end_line": true
},
{
"bbox": [
105,
598,
506,
614
],
"spans": [
{
"bbox": [
105,
598,
506,
614
],
"score": 1.0,
"content": "Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta,",
"type": "text"
}
],
"index": 37,
"is_list_start_line": true
},
{
"bbox": [
114,
609,
506,
624
],
"spans": [
{
"bbox": [
114,
609,
506,
624
],
"score": 1.0,
"content": "and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. Advances",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
116,
623,
386,
634
],
"spans": [
{
"bbox": [
116,
623,
386,
634
],
"score": 1.0,
"content": "in Neural Information Processing Systems, 33:22118–22133, 2020.",
"type": "text"
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
105,
638,
506,
654
],
"spans": [
{
"bbox": [
105,
638,
506,
654
],
"score": 1.0,
"content": "Yifan Hu. Efficient, high-quality force-directed graph drawing. Mathematica Journal, 10(1):37–71,",
"type": "text"
}
],
"index": 40,
"is_list_start_line": true
},
{
"bbox": [
114,
650,
143,
665
],
"spans": [
{
"bbox": [
114,
650,
143,
665
],
"score": 1.0,
"content": "2005.",
"type": "text"
}
],
"index": 41,
"is_list_end_line": true
},
{
"bbox": [
103,
668,
507,
684
],
"spans": [
{
"bbox": [
103,
668,
507,
684
],
"score": 1.0,
"content": "Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. ForceAtlas2, a",
"type": "text"
}
],
"index": 42,
"is_list_start_line": true
},
{
"bbox": [
115,
681,
505,
694
],
"spans": [
{
"bbox": [
115,
681,
505,
694
],
"score": 1.0,
"content": "continuous graph layout algorithm for handy network visualization designed for the gephi soft-",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
115,
692,
264,
704
],
"spans": [
{
"bbox": [
115,
692,
264,
704
],
"score": 1.0,
"content": "ware. PloS One, 9(6):e98679, 2014.",
"type": "text"
}
],
"index": 44,
"is_list_end_line": true
},
{
"bbox": [
104,
708,
506,
724
],
"spans": [
{
"bbox": [
104,
708,
506,
724
],
"score": 1.0,
"content": "Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the",
"type": "text"
}
],
"index": 45,
"is_list_start_line": true
},
{
"bbox": [
115,
720,
398,
735
],
"spans": [
{
"bbox": [
115,
720,
398,
735
],
"score": 1.0,
"content": "limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.",
"type": "text"
}
],
"index": 46,
"is_list_end_line": true
},
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
114,
92,
310,
108
],
"spans": [
{
"bbox": [
114,
92,
310,
108
],
"score": 1.0,
"content": "Conference on Learning Representations, 2015.",
"type": "text",
"cross_page": true
}
],
"index": 1,
"is_list_end_line": true
},
{
"bbox": [
105,
112,
506,
127
],
"spans": [
{
"bbox": [
105,
112,
506,
127
],
"score": 1.0,
"content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_start_line": true
},
{
"bbox": [
114,
124,
398,
138
],
"spans": [
{
"bbox": [
114,
124,
398,
138
],
"score": 1.0,
"content": "works. International Conference for Learning Representations, 2017.",
"type": "text",
"cross_page": true
}
],
"index": 3,
"is_list_end_line": true
},
{
"bbox": [
105,
144,
506,
159
],
"spans": [
{
"bbox": [
105,
144,
506,
159
],
"score": 1.0,
"content": "Dmitry Kobak and Philipp Berens. The art of using t-SNE for single-cell transcriptomics. Nature",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_start_line": true
},
{
"bbox": [
114,
155,
265,
169
],
"spans": [
{
"bbox": [
114,
155,
265,
169
],
"score": 1.0,
"content": "Communications, 10(1):5416, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_end_line": true
},
{
"bbox": [
105,
176,
506,
190
],
"spans": [
{
"bbox": [
105,
176,
506,
190
],
"score": 1.0,
"content": "Dmitry Kobak and George C Linderman. Initialization is critical for preserving global data structure",
"type": "text",
"cross_page": true
}
],
"index": 6,
"is_list_start_line": true
},
{
"bbox": [
115,
187,
407,
200
],
"spans": [
{
"bbox": [
115,
187,
407,
200
],
"score": 1.0,
"content": "in both t-SNE and UMAP. Nature Biotechnology, 39(2):156–157, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 7,
"is_list_end_line": true
},
{
"bbox": [
105,
205,
506,
221
],
"spans": [
{
"bbox": [
105,
205,
506,
221
],
"score": 1.0,
"content": "Dmitry Kobak, George Linderman, Stefan Steinerberger, Yuval Kluger, and Philipp Berens. Heavy-",
"type": "text",
"cross_page": true
}
],
"index": 8,
"is_list_start_line": true
},
{
"bbox": [
114,
218,
506,
232
],
"spans": [
{
"bbox": [
114,
218,
506,
232
],
"score": 1.0,
"content": "tailed kernels reveal a finer cluster structure in t-SNE visualisations. In Joint European Conference",
"type": "text",
"cross_page": true
}
],
"index": 9
},
{
"bbox": [
113,
227,
489,
243
],
"spans": [
{
"bbox": [
113,
227,
489,
243
],
"score": 1.0,
"content": "on Machine Learning and Knowledge Discovery in Databases, pp. 124–139. Springer, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 10,
"is_list_end_line": true
},
{
"bbox": [
105,
247,
506,
263
],
"spans": [
{
"bbox": [
105,
247,
506,
263
],
"score": 1.0,
"content": "Johannes F Kruiger, Paulo E Rauber, Rafael Messias Martins, Andreas Kerren, Stephen Kobourov,",
"type": "text",
"cross_page": true
}
],
"index": 11,
"is_list_start_line": true
},
{
"bbox": [
114,
258,
506,
276
],
"spans": [
{
"bbox": [
114,
258,
506,
276
],
"score": 1.0,
"content": "and Alexandru C Telea. Graph layouts by t-SNE. In Computer Graphics Forum, volume 36, pp.",
"type": "text",
"cross_page": true
}
],
"index": 12
},
{
"bbox": [
115,
271,
272,
285
],
"spans": [
{
"bbox": [
115,
271,
272,
285
],
"score": 1.0,
"content": "283–294. Wiley Online Library, 2017.",
"type": "text",
"cross_page": true
}
],
"index": 13,
"is_list_end_line": true
},
{
"bbox": [
105,
289,
506,
305
],
"spans": [
{
"bbox": [
105,
289,
506,
305
],
"score": 1.0,
"content": "Namkyeong Lee, Junseok Lee, and Chanyoung Park. Augmentation-free self-supervised learning",
"type": "text",
"cross_page": true
}
],
"index": 14,
"is_list_start_line": true
},
{
"bbox": [
114,
300,
505,
316
],
"spans": [
{
"bbox": [
114,
300,
505,
316
],
"score": 1.0,
"content": "on graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.",
"type": "text",
"cross_page": true
}
],
"index": 15
},
{
"bbox": [
115,
313,
191,
325
],
"spans": [
{
"bbox": [
115,
313,
191,
325
],
"score": 1.0,
"content": "7372–7380, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 16,
"is_list_end_line": true
},
{
"bbox": [
105,
331,
506,
347
],
"spans": [
{
"bbox": [
105,
331,
506,
347
],
"score": 1.0,
"content": "Yao Yang Leow, Thomas Laurent, and Xavier Bresson. GraphTSNE: a visualization technique",
"type": "text",
"cross_page": true
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
114,
344,
506,
358
],
"spans": [
{
"bbox": [
114,
344,
506,
358
],
"score": 1.0,
"content": "for graph-structured data. Representation Learning on Graphs and Manifold Workshop at the",
"type": "text",
"cross_page": true
}
],
"index": 18
},
{
"bbox": [
114,
355,
366,
369
],
"spans": [
{
"bbox": [
114,
355,
366,
369
],
"score": 1.0,
"content": "International Conference for Learning Representations, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
105,
375,
506,
389
],
"spans": [
{
"bbox": [
105,
375,
506,
389
],
"score": 1.0,
"content": "Haifeng Li, Jun Cao, Jiawei Zhu, Qinyao Luo, Silu He, and Xuying Wang. Augmentation-free graph",
"type": "text",
"cross_page": true
}
],
"index": 20,
"is_list_start_line": true
},
{
"bbox": [
115,
386,
506,
400
],
"spans": [
{
"bbox": [
115,
386,
506,
400
],
"score": 1.0,
"content": "contrastive learning of invariant-discriminative representations. IEEE Transactions on Neural",
"type": "text",
"cross_page": true
}
],
"index": 21
},
{
"bbox": [
114,
396,
316,
410
],
"spans": [
{
"bbox": [
114,
396,
316,
410
],
"score": 1.0,
"content": "Networks and Learning Systems, pp. 1–11, 2023.",
"type": "text",
"cross_page": true
}
],
"index": 22,
"is_list_end_line": true
},
{
"bbox": [
106,
417,
505,
429
],
"spans": [
{
"bbox": [
106,
417,
505,
429
],
"score": 1.0,
"content": "George C Linderman and Stefan Steinerberger. Clustering with t-SNE, provably. SIAM Journal on",
"type": "text",
"cross_page": true
}
],
"index": 23,
"is_list_start_line": true
},
{
"bbox": [
115,
429,
322,
441
],
"spans": [
{
"bbox": [
115,
429,
322,
441
],
"score": 1.0,
"content": "Mathematics of Data Science, 1(2):313–332, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 24,
"is_list_end_line": true
},
{
"bbox": [
105,
447,
504,
461
],
"spans": [
{
"bbox": [
105,
447,
504,
461
],
"score": 1.0,
"content": "George C Linderman, Manas Rachh, Jeremy G Hoskins, Stefan Steinerberger, and Yuval Kluger.",
"type": "text",
"cross_page": true
}
],
"index": 25,
"is_list_start_line": true
},
{
"bbox": [
114,
459,
506,
473
],
"spans": [
{
"bbox": [
114,
459,
506,
473
],
"score": 1.0,
"content": "Fast interpolation-based t-SNE for improved visualization of single-cell RNA-seq data. Nature",
"type": "text",
"cross_page": true
}
],
"index": 26
},
{
"bbox": [
115,
471,
245,
483
],
"spans": [
{
"bbox": [
115,
471,
245,
483
],
"score": 1.0,
"content": "Methods, 16(3):243–245, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 27,
"is_list_end_line": true
},
{
"bbox": [
105,
490,
506,
504
],
"spans": [
{
"bbox": [
105,
490,
506,
504
],
"score": 1.0,
"content": "Leland McInnes, John Healy, and James Melville. UMAP: Uniform manifold approximation and",
"type": "text",
"cross_page": true
}
],
"index": 28,
"is_list_start_line": true
},
{
"bbox": [
113,
500,
424,
515
],
"spans": [
{
"bbox": [
113,
500,
424,
515
],
"score": 1.0,
"content": "projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018.",
"type": "text",
"cross_page": true
}
],
"index": 29,
"is_list_end_line": true
},
{
"bbox": [
106,
522,
504,
533
],
"spans": [
{
"bbox": [
106,
522,
504,
533
],
"score": 1.0,
"content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representa-",
"type": "text",
"cross_page": true
}
],
"index": 30,
"is_list_start_line": true
},
{
"bbox": [
114,
532,
506,
546
],
"spans": [
{
"bbox": [
114,
532,
506,
546
],
"score": 1.0,
"content": "tions of words and phrases and their compositionality. Advances in Neural Information Processing",
"type": "text",
"cross_page": true
}
],
"index": 31
},
{
"bbox": [
114,
543,
194,
558
],
"spans": [
{
"bbox": [
114,
543,
194,
558
],
"score": 1.0,
"content": "Systems, 26, 2013.",
"type": "text",
"cross_page": true
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
104,
561,
506,
579
],
"spans": [
{
"bbox": [
104,
561,
506,
579
],
"score": 1.0,
"content": "Andreas Noack. Energy models for graph clustering. Journal of Graph Algorithms and Applications,",
"type": "text",
"cross_page": true
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
115,
575,
206,
587
],
"spans": [
{
"bbox": [
115,
575,
206,
587
],
"score": 1.0,
"content": "11(2):453–480, 2007.",
"type": "text",
"cross_page": true
}
],
"index": 34,
"is_list_end_line": true
},
{
"bbox": [
105,
594,
505,
608
],
"spans": [
{
"bbox": [
105,
594,
505,
608
],
"score": 1.0,
"content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-",
"type": "text",
"cross_page": true
}
],
"index": 35,
"is_list_start_line": true
},
{
"bbox": [
114,
606,
329,
618
],
"spans": [
{
"bbox": [
114,
606,
329,
618
],
"score": 1.0,
"content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.",
"type": "text",
"cross_page": true
}
],
"index": 36,
"is_list_end_line": true
},
{
"bbox": [
105,
625,
506,
639
],
"spans": [
{
"bbox": [
105,
625,
506,
639
],
"score": 1.0,
"content": "Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨",
"type": "text",
"cross_page": true
}
],
"index": 37,
"is_list_start_line": true
},
{
"bbox": [
114,
635,
506,
650
],
"spans": [
{
"bbox": [
114,
635,
506,
650
],
"score": 1.0,
"content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn:",
"type": "text",
"cross_page": true
}
],
"index": 38
},
{
"bbox": [
115,
648,
482,
660
],
"spans": [
{
"bbox": [
115,
648,
482,
660
],
"score": 1.0,
"content": "Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.",
"type": "text",
"cross_page": true
}
],
"index": 39,
"is_list_end_line": true
},
{
"bbox": [
105,
667,
506,
681
],
"spans": [
{
"bbox": [
105,
667,
506,
681
],
"score": 1.0,
"content": "Pavlin G Policar, Martin Stra ˇ zar, and Bla ˇ z Zupan. openTSNE: a modular Python library for t-SNE ˇ",
"type": "text",
"cross_page": true
}
],
"index": 40,
"is_list_start_line": true
},
{
"bbox": [
115,
678,
398,
693
],
"spans": [
{
"bbox": [
115,
678,
398,
693
],
"score": 1.0,
"content": "dimensionality reduction and embedding. BioRxiv, pp. 731877, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 41,
"is_list_end_line": true
},
{
"bbox": [
104,
696,
506,
714
],
"spans": [
{
"bbox": [
104,
696,
506,
714
],
"score": 1.0,
"content": "Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Mehdi Azabou, Eva L Dyer,",
"type": "text",
"cross_page": true
}
],
"index": 42,
"is_list_start_line": true
},
{
"bbox": [
114,
708,
506,
724
],
"spans": [
{
"bbox": [
114,
708,
506,
724
],
"score": 1.0,
"content": "Remi Munos, Petar Velickovi ˇ c, and Michal Valko. Large-scale representation learning on graphs ´",
"type": "text",
"cross_page": true
}
],
"index": 43
},
{
"bbox": [
116,
720,
354,
735
],
"spans": [
{
"bbox": [
116,
720,
354,
735
],
"score": 1.0,
"content": "via bootstrapping. arXiv preprint arXiv:2102.06514, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 44,
"is_list_end_line": true
},
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "Puja Trivedi, Ekdeep Singh Lubana, Yujun Yan, Yaoqing Yang, and Danai Koutra. Augmentations",
"type": "text",
"cross_page": true
}
],
"index": 0,
"is_list_start_line": true
},
{
"bbox": [
115,
94,
505,
108
],
"spans": [
{
"bbox": [
115,
94,
505,
108
],
"score": 1.0,
"content": "in graph contrastive learning: Current methodological flaws & towards better practices. In Pro-",
"type": "text",
"cross_page": true
}
],
"index": 1
},
{
"bbox": [
115,
105,
383,
117
],
"spans": [
{
"bbox": [
115,
105,
383,
117
],
"score": 1.0,
"content": "ceedings of the ACM Web Conference 2022, pp. 1538–1549, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 2,
"is_list_end_line": true
},
{
"bbox": [
105,
122,
505,
137
],
"spans": [
{
"bbox": [
105,
122,
505,
137
],
"score": 1.0,
"content": "Laurens Van Der Maaten. Accelerating t-SNE using tree-based algorithms. The Journal of Machine",
"type": "text",
"cross_page": true
}
],
"index": 3,
"is_list_start_line": true
},
{
"bbox": [
115,
134,
298,
148
],
"spans": [
{
"bbox": [
115,
134,
298,
148
],
"score": 1.0,
"content": "Learning Research, 15(1):3221–3245, 2014.",
"type": "text",
"cross_page": true
}
],
"index": 4,
"is_list_end_line": true
},
{
"bbox": [
104,
153,
505,
168
],
"spans": [
{
"bbox": [
104,
153,
505,
168
],
"score": 1.0,
"content": "Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine",
"type": "text",
"cross_page": true
}
],
"index": 5,
"is_list_start_line": true
},
{
"bbox": [
115,
165,
248,
177
],
"spans": [
{
"bbox": [
115,
165,
248,
177
],
"score": 1.0,
"content": "Learning Research, 9(11), 2008.",
"type": "text",
"cross_page": true
}
],
"index": 6,
"is_list_end_line": true
},
{
"bbox": [
105,
183,
505,
197
],
"spans": [
{
"bbox": [
105,
183,
505,
197
],
"score": 1.0,
"content": "Petar Velickovic, William Fedus, William L Hamilton, Pietro Lio, Yoshua Bengio, and R Devon. `",
"type": "text",
"cross_page": true
}
],
"index": 7,
"is_list_start_line": true
},
{
"bbox": [
115,
193,
331,
208
],
"spans": [
{
"bbox": [
115,
193,
331,
208
],
"score": 1.0,
"content": "676 hjelm. Deep graph infomax. ICLR, 2(3):4, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 8,
"is_list_end_line": true
},
{
"bbox": [
105,
213,
505,
227
],
"spans": [
{
"bbox": [
105,
213,
505,
227
],
"score": 1.0,
"content": "Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma,",
"type": "text",
"cross_page": true
}
],
"index": 9,
"is_list_start_line": true
},
{
"bbox": [
115,
224,
506,
238
],
"spans": [
{
"bbox": [
115,
224,
506,
238
],
"score": 1.0,
"content": "Lingfan Yu, Yu Gai, et al. Deep graph library: A graph-centric, highly-performant package for",
"type": "text",
"cross_page": true
}
],
"index": 10
},
{
"bbox": [
115,
235,
374,
248
],
"spans": [
{
"bbox": [
115,
235,
374,
248
],
"score": 1.0,
"content": "graph neural networks. arXiv preprint arXiv:1909.01315, 2019.",
"type": "text",
"cross_page": true
}
],
"index": 11,
"is_list_end_line": true
},
{
"bbox": [
105,
253,
505,
268
],
"spans": [
{
"bbox": [
105,
253,
505,
268
],
"score": 1.0,
"content": "Yifei Wang, Qi Zhang, Tianqi Du, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. A message",
"type": "text",
"cross_page": true
}
],
"index": 12,
"is_list_start_line": true
},
{
"bbox": [
115,
265,
506,
279
],
"spans": [
{
"bbox": [
115,
265,
506,
279
],
"score": 1.0,
"content": "passing perspective on learning dynamics of contrastive learning. International Conference on",
"type": "text",
"cross_page": true
}
],
"index": 13
},
{
"bbox": [
115,
276,
250,
290
],
"spans": [
{
"bbox": [
115,
276,
250,
290
],
"score": 1.0,
"content": "Learning Representations, 2023.",
"type": "text",
"cross_page": true
}
],
"index": 14,
"is_list_end_line": true
},
{
"bbox": [
105,
294,
505,
309
],
"spans": [
{
"bbox": [
105,
294,
505,
309
],
"score": 1.0,
"content": "Zhirong Yang, Jaakko Peltonen, and Samuel Kaski. Scalable optimization of neighbor embedding",
"type": "text",
"cross_page": true
}
],
"index": 15,
"is_list_start_line": true
},
{
"bbox": [
115,
306,
505,
320
],
"spans": [
{
"bbox": [
115,
306,
505,
320
],
"score": 1.0,
"content": "for visualization. In International Conference on Machine Learning, pp. 127–135. PMLR, 2013.",
"type": "text",
"cross_page": true
}
],
"index": 16
},
{
"bbox": [
105,
324,
505,
339
],
"spans": [
{
"bbox": [
105,
324,
505,
339
],
"score": 1.0,
"content": "Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph",
"type": "text",
"cross_page": true
}
],
"index": 17,
"is_list_start_line": true
},
{
"bbox": [
115,
336,
506,
350
],
"spans": [
{
"bbox": [
115,
336,
506,
350
],
"score": 1.0,
"content": "contrastive learning with augmentations. Advances in Neural Information Processing Systems,",
"type": "text",
"cross_page": true
}
],
"index": 18
},
{
"bbox": [
117,
347,
203,
358
],
"spans": [
{
"bbox": [
117,
347,
203,
358
],
"score": 1.0,
"content": "33:5812–5823, 2020.",
"type": "text",
"cross_page": true
}
],
"index": 19,
"is_list_end_line": true
},
{
"bbox": [
106,
366,
505,
378
],
"spans": [
{
"bbox": [
106,
366,
505,
378
],
"score": 1.0,
"content": "Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. From canonical correlation",
"type": "text",
"cross_page": true
}
],
"index": 20,
"is_list_start_line": true
},
{
"bbox": [
114,
375,
506,
392
],
"spans": [
{
"bbox": [
114,
375,
506,
392
],
"score": 1.0,
"content": "analysis to self-supervised graph neural networks. Advances in Neural Information Processing",
"type": "text",
"cross_page": true
}
],
"index": 21
},
{
"bbox": [
115,
387,
222,
401
],
"spans": [
{
"bbox": [
115,
387,
222,
401
],
"score": 1.0,
"content": "Systems, 34:76–89, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 22,
"is_list_end_line": true
},
{
"bbox": [
105,
406,
506,
420
],
"spans": [
{
"bbox": [
105,
406,
506,
420
],
"score": 1.0,
"content": "Hengrui Zhang, Qitian Wu, Yu Wang, Shaofeng Zhang, Junchi Yan, and Philip S Yu. Localized",
"type": "text",
"cross_page": true
}
],
"index": 23,
"is_list_start_line": true
},
{
"bbox": [
115,
417,
404,
431
],
"spans": [
{
"bbox": [
115,
417,
404,
431
],
"score": 1.0,
"content": "contrastive learning on graphs. arXiv preprint arXiv:2212.04604, 2022.",
"type": "text",
"cross_page": true
}
],
"index": 24,
"is_list_end_line": true
},
{
"bbox": [
105,
434,
505,
450
],
"spans": [
{
"bbox": [
105,
434,
505,
450
],
"score": 1.0,
"content": "Fahai Zhong, Mingliang Xue, Jian Zhang, Fan Zhang, Rui Ban, Oliver Deussen, and Yunhai Wang.",
"type": "text",
"cross_page": true
}
],
"index": 25,
"is_list_start_line": true
},
{
"bbox": [
115,
447,
505,
461
],
"spans": [
{
"bbox": [
115,
447,
505,
461
],
"score": 1.0,
"content": "Force-directed graph layouts revisited: a new force based on the t-distribution. IEEE Transactions",
"type": "text",
"cross_page": true
}
],
"index": 26
},
{
"bbox": [
115,
459,
310,
471
],
"spans": [
{
"bbox": [
115,
459,
310,
471
],
"score": 1.0,
"content": "on Visualization and Computer Graphics, 2023.",
"type": "text",
"cross_page": true
}
],
"index": 27,
"is_list_end_line": true
},
{
"bbox": [
105,
477,
505,
491
],
"spans": [
{
"bbox": [
105,
477,
505,
491
],
"score": 1.0,
"content": "Minfeng Zhu, Wei Chen, Yuanzhe Hu, Yuxuan Hou, Liangjun Liu, and Kaiyuan Zhang. DRGraph:",
"type": "text",
"cross_page": true
}
],
"index": 28,
"is_list_start_line": true
},
{
"bbox": [
115,
487,
507,
503
],
"spans": [
{
"bbox": [
115,
487,
507,
503
],
"score": 1.0,
"content": "An efficient graph layout algorithm for large-scale graphs by dimensionality reduction. IEEE",
"type": "text",
"cross_page": true
}
],
"index": 29
},
{
"bbox": [
115,
497,
442,
513
],
"spans": [
{
"bbox": [
115,
497,
442,
513
],
"score": 1.0,
"content": "Transactions on Visualization and Computer Graphics, 27(2):1666–1676, 2020a.",
"type": "text",
"cross_page": true
}
],
"index": 30,
"is_list_end_line": true
},
{
"bbox": [
105,
517,
505,
532
],
"spans": [
{
"bbox": [
105,
517,
505,
532
],
"score": 1.0,
"content": "Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive",
"type": "text",
"cross_page": true
}
],
"index": 31,
"is_list_start_line": true
},
{
"bbox": [
114,
529,
381,
542
],
"spans": [
{
"bbox": [
114,
529,
381,
542
],
"score": 1.0,
"content": "representation learning. arXiv preprint arXiv:2006.04131, 2020b.",
"type": "text",
"cross_page": true
}
],
"index": 32,
"is_list_end_line": true
},
{
"bbox": [
106,
546,
505,
563
],
"spans": [
{
"bbox": [
106,
546,
505,
563
],
"score": 1.0,
"content": "Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning",
"type": "text",
"cross_page": true
}
],
"index": 33,
"is_list_start_line": true
},
{
"bbox": [
115,
559,
503,
573
],
"spans": [
{
"bbox": [
115,
559,
503,
573
],
"score": 1.0,
"content": "with adaptive augmentation. In Proceedings of the Web Conference 2021, pp. 2069–2080, 2021.",
"type": "text",
"cross_page": true
}
],
"index": 34
}
],
"index": 24.5,
"bbox_fs": [
103,
129,
507,
735
]
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
104,
51,
506,
739
],
"lines": [
{
"bbox": [
105,
81,
506,
96
],
"spans": [
{
"bbox": [
105,
81,
506,
96
],
"score": 1.0,
"content": "Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
114,
92,
310,
108
],
"spans": [
{
"bbox": [
114,
92,
310,
108
],
"score": 1.0,
"content": "Conference on Learning Representations, 2015.",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
105,
112,
506,
127
],
"spans": [
{
"bbox": [
105,
112,
506,
127
],
"score": 1.0,
"content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
114,
124,
398,
138
],
"spans": [
{
"bbox": [
114,
124,
398,
138
],
"score": 1.0,
"content": "works. International Conference for Learning Representations, 2017.",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
105,
144,
506,
159
],
"spans": [
{
"bbox": [
105,
144,
506,
159
],
"score": 1.0,
"content": "Dmitry Kobak and Philipp Berens. The art of using t-SNE for single-cell transcriptomics. Nature",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
114,
155,
265,
169
],
"spans": [
{
"bbox": [
114,
155,
265,
169
],
"score": 1.0,
"content": "Communications, 10(1):5416, 2019.",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
105,
176,
506,
190
],
"spans": [
{
"bbox": [
105,
176,
506,
190
],
"score": 1.0,
"content": "Dmitry Kobak and George C Linderman. Initialization is critical for preserving global data structure",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
115,
187,
407,
200
],
"spans": [
{
"bbox": [
115,
187,
407,
200
],
"score": 1.0,
"content": "in both t-SNE and UMAP. Nature Biotechnology, 39(2):156–157, 2021.",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
105,
205,
506,
221
],
"spans": [
{
"bbox": [
105,
205,
506,
221
],
"score": 1.0,
"content": "Dmitry Kobak, George Linderman, Stefan Steinerberger, Yuval Kluger, and Philipp Berens. Heavy-",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
114,
218,
506,
232
],
"spans": [
{
"bbox": [
114,
218,
506,
232
],
"score": 1.0,
"content": "tailed kernels reveal a finer cluster structure in t-SNE visualisations. In Joint European Conference",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
113,
227,
489,
243
],
"spans": [
{
"bbox": [
113,
227,
489,
243
],
"score": 1.0,
"content": "on Machine Learning and Knowledge Discovery in Databases, pp. 124–139. Springer, 2019.",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
105,
247,
506,
263
],
"spans": [
{
"bbox": [
105,
247,
506,
263
],
"score": 1.0,
"content": "Johannes F Kruiger, Paulo E Rauber, Rafael Messias Martins, Andreas Kerren, Stephen Kobourov,",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
114,
258,
506,
276
],
"spans": [
{
"bbox": [
114,
258,
506,
276
],
"score": 1.0,
"content": "and Alexandru C Telea. Graph layouts by t-SNE. In Computer Graphics Forum, volume 36, pp.",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
115,
271,
272,
285
],
"spans": [
{
"bbox": [
115,
271,
272,
285
],
"score": 1.0,
"content": "283–294. Wiley Online Library, 2017.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
289,
506,
305
],
"spans": [
{
"bbox": [
105,
289,
506,
305
],
"score": 1.0,
"content": "Namkyeong Lee, Junseok Lee, and Chanyoung Park. Augmentation-free self-supervised learning",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
114,
300,
505,
316
],
"spans": [
{
"bbox": [
114,
300,
505,
316
],
"score": 1.0,
"content": "on graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
313,
191,
325
],
"spans": [
{
"bbox": [
115,
313,
191,
325
],
"score": 1.0,
"content": "7372–7380, 2022.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
331,
506,
347
],
"spans": [
{
"bbox": [
105,
331,
506,
347
],
"score": 1.0,
"content": "Yao Yang Leow, Thomas Laurent, and Xavier Bresson. GraphTSNE: a visualization technique",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
114,
344,
506,
358
],
"spans": [
{
"bbox": [
114,
344,
506,
358
],
"score": 1.0,
"content": "for graph-structured data. Representation Learning on Graphs and Manifold Workshop at the",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
114,
355,
366,
369
],
"spans": [
{
"bbox": [
114,
355,
366,
369
],
"score": 1.0,
"content": "International Conference for Learning Representations, 2019.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
105,
375,
506,
389
],
"spans": [
{
"bbox": [
105,
375,
506,
389
],
"score": 1.0,
"content": "Haifeng Li, Jun Cao, Jiawei Zhu, Qinyao Luo, Silu He, and Xuying Wang. Augmentation-free graph",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
115,
386,
506,
400
],
"spans": [
{
"bbox": [
115,
386,
506,
400
],
"score": 1.0,
"content": "contrastive learning of invariant-discriminative representations. IEEE Transactions on Neural",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
114,
396,
316,
410
],
"spans": [
{
"bbox": [
114,
396,
316,
410
],
"score": 1.0,
"content": "Networks and Learning Systems, pp. 1–11, 2023.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
106,
417,
505,
429
],
"spans": [
{
"bbox": [
106,
417,
505,
429
],
"score": 1.0,
"content": "George C Linderman and Stefan Steinerberger. Clustering with t-SNE, provably. SIAM Journal on",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
429,
322,
441
],
"spans": [
{
"bbox": [
115,
429,
322,
441
],
"score": 1.0,
"content": "Mathematics of Data Science, 1(2):313–332, 2019.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
447,
504,
461
],
"spans": [
{
"bbox": [
105,
447,
504,
461
],
"score": 1.0,
"content": "George C Linderman, Manas Rachh, Jeremy G Hoskins, Stefan Steinerberger, and Yuval Kluger.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
114,
459,
506,
473
],
"spans": [
{
"bbox": [
114,
459,
506,
473
],
"score": 1.0,
"content": "Fast interpolation-based t-SNE for improved visualization of single-cell RNA-seq data. Nature",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
471,
245,
483
],
"spans": [
{
"bbox": [
115,
471,
245,
483
],
"score": 1.0,
"content": "Methods, 16(3):243–245, 2019.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
490,
506,
504
],
"spans": [
{
"bbox": [
105,
490,
506,
504
],
"score": 1.0,
"content": "Leland McInnes, John Healy, and James Melville. UMAP: Uniform manifold approximation and",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
113,
500,
424,
515
],
"spans": [
{
"bbox": [
113,
500,
424,
515
],
"score": 1.0,
"content": "projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018.",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
106,
522,
504,
533
],
"spans": [
{
"bbox": [
106,
522,
504,
533
],
"score": 1.0,
"content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representa-",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
114,
532,
506,
546
],
"spans": [
{
"bbox": [
114,
532,
506,
546
],
"score": 1.0,
"content": "tions of words and phrases and their compositionality. Advances in Neural Information Processing",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
114,
543,
194,
558
],
"spans": [
{
"bbox": [
114,
543,
194,
558
],
"score": 1.0,
"content": "Systems, 26, 2013.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
104,
561,
506,
579
],
"spans": [
{
"bbox": [
104,
561,
506,
579
],
"score": 1.0,
"content": "Andreas Noack. Energy models for graph clustering. Journal of Graph Algorithms and Applications,",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
575,
206,
587
],
"spans": [
{
"bbox": [
115,
575,
206,
587
],
"score": 1.0,
"content": "11(2):453–480, 2007.",
"type": "text"
}
],
"index": 34
},
{
"bbox": [
105,
594,
505,
608
],
"spans": [
{
"bbox": [
105,
594,
505,
608
],
"score": 1.0,
"content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-",
"type": "text"
}
],
"index": 35
},
{
"bbox": [
114,
606,
329,
618
],
"spans": [
{
"bbox": [
114,
606,
329,
618
],
"score": 1.0,
"content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.",
"type": "text"
}
],
"index": 36
},
{
"bbox": [
105,
625,
506,
639
],
"spans": [
{
"bbox": [
105,
625,
506,
639
],
"score": 1.0,
"content": "Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier ¨",
"type": "text"
}
],
"index": 37
},
{
"bbox": [
114,
635,
506,
650
],
"spans": [
{
"bbox": [
114,
635,
506,
650
],
"score": 1.0,
"content": "Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn:",
"type": "text"
}
],
"index": 38
},
{
"bbox": [
115,
648,
482,
660
],
"spans": [
{
"bbox": [
115,
648,
482,
660
],
"score": 1.0,
"content": "Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.",
"type": "text"
}
],
"index": 39
},
{
"bbox": [
105,
667,
506,
681
],
"spans": [
{
"bbox": [
105,
667,
506,
681
],
"score": 1.0,
"content": "Pavlin G Policar, Martin Stra ˇ zar, and Bla ˇ z Zupan. openTSNE: a modular Python library for t-SNE ˇ",
"type": "text"
}
],
"index": 40
},
{
"bbox": [
115,
678,
398,
693
],
"spans": [
{
"bbox": [
115,
678,
398,
693
],
"score": 1.0,
"content": "dimensionality reduction and embedding. BioRxiv, pp. 731877, 2019.",
"type": "text"
}
],
"index": 41
},
{
"bbox": [
104,
696,
506,
714
],
"spans": [
{
"bbox": [
104,
696,
506,
714
],
"score": 1.0,
"content": "Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Mehdi Azabou, Eva L Dyer,",
"type": "text"
}
],
"index": 42
},
{
"bbox": [
114,
708,
506,
724
],
"spans": [
{
"bbox": [
114,
708,
506,
724
],
"score": 1.0,
"content": "Remi Munos, Petar Velickovi ˇ c, and Michal Valko. Large-scale representation learning on graphs ´",
"type": "text"
}
],
"index": 43
},
{
"bbox": [
116,
720,
354,
735
],
"spans": [
{
"bbox": [
116,
720,
354,
735
],
"score": 1.0,
"content": "via bootstrapping. arXiv preprint arXiv:2102.06514, 2021.",
"type": "text"
}
],
"index": 44
}
],
"index": 22
}
],
"page_idx": 10,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
310,
760
],
"lines": [
{
"bbox": [
299,
750,
312,
765
],
"spans": [
{
"bbox": [
299,
750,
312,
765
],
"score": 1.0,
"content": "",
"type": "text",
"height": 15,
"width": 13
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
104,
51,
506,
739
],
"lines": [],
"index": 22,
"bbox_fs": [
104,
81,
506,
735
],
"lines_deleted": true
}
]
},
{
"preproc_blocks": [
{
"type": "text",
"bbox": [
104,
72,
506,
572
],
"lines": [
{
"bbox": [
106,
83,
505,
95
],
"spans": [
{
"bbox": [
106,
83,
505,
95
],
"score": 1.0,
"content": "Puja Trivedi, Ekdeep Singh Lubana, Yujun Yan, Yaoqing Yang, and Danai Koutra. Augmentations",
"type": "text"
}
],
"index": 0
},
{
"bbox": [
115,
94,
505,
108
],
"spans": [
{
"bbox": [
115,
94,
505,
108
],
"score": 1.0,
"content": "in graph contrastive learning: Current methodological flaws & towards better practices. In Pro-",
"type": "text"
}
],
"index": 1
},
{
"bbox": [
115,
105,
383,
117
],
"spans": [
{
"bbox": [
115,
105,
383,
117
],
"score": 1.0,
"content": "ceedings of the ACM Web Conference 2022, pp. 1538–1549, 2022.",
"type": "text"
}
],
"index": 2
},
{
"bbox": [
105,
122,
505,
137
],
"spans": [
{
"bbox": [
105,
122,
505,
137
],
"score": 1.0,
"content": "Laurens Van Der Maaten. Accelerating t-SNE using tree-based algorithms. The Journal of Machine",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
115,
134,
298,
148
],
"spans": [
{
"bbox": [
115,
134,
298,
148
],
"score": 1.0,
"content": "Learning Research, 15(1):3221–3245, 2014.",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
104,
153,
505,
168
],
"spans": [
{
"bbox": [
104,
153,
505,
168
],
"score": 1.0,
"content": "Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine",
"type": "text"
}
],
"index": 5
},
{
"bbox": [
115,
165,
248,
177
],
"spans": [
{
"bbox": [
115,
165,
248,
177
],
"score": 1.0,
"content": "Learning Research, 9(11), 2008.",
"type": "text"
}
],
"index": 6
},
{
"bbox": [
105,
183,
505,
197
],
"spans": [
{
"bbox": [
105,
183,
505,
197
],
"score": 1.0,
"content": "Petar Velickovic, William Fedus, William L Hamilton, Pietro Lio, Yoshua Bengio, and R Devon. `",
"type": "text"
}
],
"index": 7
},
{
"bbox": [
115,
193,
331,
208
],
"spans": [
{
"bbox": [
115,
193,
331,
208
],
"score": 1.0,
"content": "676 hjelm. Deep graph infomax. ICLR, 2(3):4, 2019.",
"type": "text"
}
],
"index": 8
},
{
"bbox": [
105,
213,
505,
227
],
"spans": [
{
"bbox": [
105,
213,
505,
227
],
"score": 1.0,
"content": "Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma,",
"type": "text"
}
],
"index": 9
},
{
"bbox": [
115,
224,
506,
238
],
"spans": [
{
"bbox": [
115,
224,
506,
238
],
"score": 1.0,
"content": "Lingfan Yu, Yu Gai, et al. Deep graph library: A graph-centric, highly-performant package for",
"type": "text"
}
],
"index": 10
},
{
"bbox": [
115,
235,
374,
248
],
"spans": [
{
"bbox": [
115,
235,
374,
248
],
"score": 1.0,
"content": "graph neural networks. arXiv preprint arXiv:1909.01315, 2019.",
"type": "text"
}
],
"index": 11
},
{
"bbox": [
105,
253,
505,
268
],
"spans": [
{
"bbox": [
105,
253,
505,
268
],
"score": 1.0,
"content": "Yifei Wang, Qi Zhang, Tianqi Du, Jiansheng Yang, Zhouchen Lin, and Yisen Wang. A message",
"type": "text"
}
],
"index": 12
},
{
"bbox": [
115,
265,
506,
279
],
"spans": [
{
"bbox": [
115,
265,
506,
279
],
"score": 1.0,
"content": "passing perspective on learning dynamics of contrastive learning. International Conference on",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
115,
276,
250,
290
],
"spans": [
{
"bbox": [
115,
276,
250,
290
],
"score": 1.0,
"content": "Learning Representations, 2023.",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
294,
505,
309
],
"spans": [
{
"bbox": [
105,
294,
505,
309
],
"score": 1.0,
"content": "Zhirong Yang, Jaakko Peltonen, and Samuel Kaski. Scalable optimization of neighbor embedding",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
115,
306,
505,
320
],
"spans": [
{
"bbox": [
115,
306,
505,
320
],
"score": 1.0,
"content": "for visualization. In International Conference on Machine Learning, pp. 127–135. PMLR, 2013.",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
324,
505,
339
],
"spans": [
{
"bbox": [
105,
324,
505,
339
],
"score": 1.0,
"content": "Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
115,
336,
506,
350
],
"spans": [
{
"bbox": [
115,
336,
506,
350
],
"score": 1.0,
"content": "contrastive learning with augmentations. Advances in Neural Information Processing Systems,",
"type": "text"
}
],
"index": 18
},
{
"bbox": [
117,
347,
203,
358
],
"spans": [
{
"bbox": [
117,
347,
203,
358
],
"score": 1.0,
"content": "33:5812–5823, 2020.",
"type": "text"
}
],
"index": 19
},
{
"bbox": [
106,
366,
505,
378
],
"spans": [
{
"bbox": [
106,
366,
505,
378
],
"score": 1.0,
"content": "Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. From canonical correlation",
"type": "text"
}
],
"index": 20
},
{
"bbox": [
114,
375,
506,
392
],
"spans": [
{
"bbox": [
114,
375,
506,
392
],
"score": 1.0,
"content": "analysis to self-supervised graph neural networks. Advances in Neural Information Processing",
"type": "text"
}
],
"index": 21
},
{
"bbox": [
115,
387,
222,
401
],
"spans": [
{
"bbox": [
115,
387,
222,
401
],
"score": 1.0,
"content": "Systems, 34:76–89, 2021.",
"type": "text"
}
],
"index": 22
},
{
"bbox": [
105,
406,
506,
420
],
"spans": [
{
"bbox": [
105,
406,
506,
420
],
"score": 1.0,
"content": "Hengrui Zhang, Qitian Wu, Yu Wang, Shaofeng Zhang, Junchi Yan, and Philip S Yu. Localized",
"type": "text"
}
],
"index": 23
},
{
"bbox": [
115,
417,
404,
431
],
"spans": [
{
"bbox": [
115,
417,
404,
431
],
"score": 1.0,
"content": "contrastive learning on graphs. arXiv preprint arXiv:2212.04604, 2022.",
"type": "text"
}
],
"index": 24
},
{
"bbox": [
105,
434,
505,
450
],
"spans": [
{
"bbox": [
105,
434,
505,
450
],
"score": 1.0,
"content": "Fahai Zhong, Mingliang Xue, Jian Zhang, Fan Zhang, Rui Ban, Oliver Deussen, and Yunhai Wang.",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
115,
447,
505,
461
],
"spans": [
{
"bbox": [
115,
447,
505,
461
],
"score": 1.0,
"content": "Force-directed graph layouts revisited: a new force based on the t-distribution. IEEE Transactions",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
115,
459,
310,
471
],
"spans": [
{
"bbox": [
115,
459,
310,
471
],
"score": 1.0,
"content": "on Visualization and Computer Graphics, 2023.",
"type": "text"
}
],
"index": 27
},
{
"bbox": [
105,
477,
505,
491
],
"spans": [
{
"bbox": [
105,
477,
505,
491
],
"score": 1.0,
"content": "Minfeng Zhu, Wei Chen, Yuanzhe Hu, Yuxuan Hou, Liangjun Liu, and Kaiyuan Zhang. DRGraph:",
"type": "text"
}
],
"index": 28
},
{
"bbox": [
115,
487,
507,
503
],
"spans": [
{
"bbox": [
115,
487,
507,
503
],
"score": 1.0,
"content": "An efficient graph layout algorithm for large-scale graphs by dimensionality reduction. IEEE",
"type": "text"
}
],
"index": 29
},
{
"bbox": [
115,
497,
442,
513
],
"spans": [
{
"bbox": [
115,
497,
442,
513
],
"score": 1.0,
"content": "Transactions on Visualization and Computer Graphics, 27(2):1666–1676, 2020a.",
"type": "text"
}
],
"index": 30
},
{
"bbox": [
105,
517,
505,
532
],
"spans": [
{
"bbox": [
105,
517,
505,
532
],
"score": 1.0,
"content": "Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Deep graph contrastive",
"type": "text"
}
],
"index": 31
},
{
"bbox": [
114,
529,
381,
542
],
"spans": [
{
"bbox": [
114,
529,
381,
542
],
"score": 1.0,
"content": "representation learning. arXiv preprint arXiv:2006.04131, 2020b.",
"type": "text"
}
],
"index": 32
},
{
"bbox": [
106,
546,
505,
563
],
"spans": [
{
"bbox": [
106,
546,
505,
563
],
"score": 1.0,
"content": "Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. Graph contrastive learning",
"type": "text"
}
],
"index": 33
},
{
"bbox": [
115,
559,
503,
573
],
"spans": [
{
"bbox": [
115,
559,
503,
573
],
"score": 1.0,
"content": "with adaptive augmentation. In Proceedings of the Web Conference 2021, pp. 2069–2080, 2021.",
"type": "text"
}
],
"index": 34
}
],
"index": 17
}
],
"page_idx": 11,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
107,
27,
308,
37
],
"lines": [
{
"bbox": [
107,
26,
308,
38
],
"spans": [
{
"bbox": [
107,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
313,
764
],
"spans": [
{
"bbox": [
299,
750,
313,
764
],
"score": 1.0,
"content": "12",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "list",
"bbox": [
104,
72,
506,
572
],
"lines": [],
"index": 17,
"bbox_fs": [
104,
83,
507,
573
],
"lines_deleted": true
}
]
},
{
"preproc_blocks": [
{
"type": "title",
"bbox": [
108,
81,
266,
94
],
"lines": [
{
"bbox": [
106,
80,
266,
96
],
"spans": [
{
"bbox": [
106,
80,
266,
96
],
"score": 1.0,
"content": "A SUPPLEMENTARY FIGURES",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image",
"bbox": [
216,
111,
395,
268
],
"blocks": [
{
"type": "image_body",
"bbox": [
216,
111,
395,
268
],
"group_id": 0,
"lines": [
{
"bbox": [
216,
111,
395,
268
],
"spans": [
{
"bbox": [
216,
111,
395,
268
],
"score": 0.968,
"type": "image",
"image_path": "60b7d14ab4a629377a795bcc99485453a195556eaadd495831ef6798a0248ae6.jpg"
}
]
}
],
"index": 6.5,
"virtual_lines": [
{
"bbox": [
216,
111,
395,
124.08333333333333
],
"spans": [],
"index": 1
},
{
"bbox": [
216,
124.08333333333333,
395,
137.16666666666666
],
"spans": [],
"index": 2
},
{
"bbox": [
216,
137.16666666666666,
395,
150.25
],
"spans": [],
"index": 3
},
{
"bbox": [
216,
150.25,
395,
163.33333333333334
],
"spans": [],
"index": 4
},
{
"bbox": [
216,
163.33333333333334,
395,
176.41666666666669
],
"spans": [],
"index": 5
},
{
"bbox": [
216,
176.41666666666669,
395,
189.50000000000003
],
"spans": [],
"index": 6
},
{
"bbox": [
216,
189.50000000000003,
395,
202.58333333333337
],
"spans": [],
"index": 7
},
{
"bbox": [
216,
202.58333333333337,
395,
215.6666666666667
],
"spans": [],
"index": 8
},
{
"bbox": [
216,
215.6666666666667,
395,
228.75000000000006
],
"spans": [],
"index": 9
},
{
"bbox": [
216,
228.75000000000006,
395,
241.8333333333334
],
"spans": [],
"index": 10
},
{
"bbox": [
216,
241.8333333333334,
395,
254.91666666666674
],
"spans": [],
"index": 11
},
{
"bbox": [
216,
254.91666666666674,
395,
268.00000000000006
],
"spans": [],
"index": 12
}
]
},
{
"type": "image_caption",
"bbox": [
106,
286,
505,
354
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
287,
505,
300
],
"spans": [
{
"bbox": [
106,
287,
272,
300
],
"score": 1.0,
"content": "Figure A.1: Computation times for graph",
"type": "text"
},
{
"bbox": [
273,
288,
278,
298
],
"score": 0.61,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
287,
505,
300
],
"score": 1.0,
"content": "-SNE and graph CNE with 2 and 128 output dimensions.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
298,
505,
311
],
"spans": [
{
"bbox": [
105,
298,
242,
311
],
"score": 1.0,
"content": "openTSNE was run on CPU with",
"type": "text"
},
{
"bbox": [
243,
299,
294,
309
],
"score": 0.85,
"content": "\\mathrm { n - j } \\mathrm { o b s = - 1 }",
"type": "inline_equation"
},
{
"bbox": [
294,
298,
505,
311
],
"score": 1.0,
"content": ". CNE was run on GPU. Datasets are ordered by the",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
309,
505,
322
],
"spans": [
{
"bbox": [
105,
309,
505,
322
],
"score": 1.0,
"content": "increasing number of nodes. The runtime of openTSNE (for a given number of gradient descent",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
320,
506,
333
],
"spans": [
{
"bbox": [
105,
320,
506,
333
],
"score": 1.0,
"content": "steps) grows linearly with the number of nodes. The runtime of CNE (for a given number of epochs",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
331,
505,
344
],
"spans": [
{
"bbox": [
105,
331,
505,
344
],
"score": 1.0,
"content": "and a given batch size) grows linearly with the number of edges. The Pubmed dataset (PUB) has",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
342,
383,
354
],
"spans": [
{
"bbox": [
105,
342,
383,
354
],
"score": 1.0,
"content": "fewer edges than the Amazon datasets (APH and ACO), see Table 1.",
"type": "text"
}
],
"index": 18
}
],
"index": 15.5
}
],
"index": 11.0
}
],
"page_idx": 12,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
313,
764
],
"spans": [
{
"bbox": [
299,
750,
313,
764
],
"score": 1.0,
"content": "13",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
25,
308,
38
],
"spans": [
{
"bbox": [
106,
25,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "title",
"bbox": [
108,
81,
266,
94
],
"lines": [
{
"bbox": [
106,
80,
266,
96
],
"spans": [
{
"bbox": [
106,
80,
266,
96
],
"score": 1.0,
"content": "A SUPPLEMENTARY FIGURES",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image",
"bbox": [
216,
111,
395,
268
],
"blocks": [
{
"type": "image_body",
"bbox": [
216,
111,
395,
268
],
"group_id": 0,
"lines": [
{
"bbox": [
216,
111,
395,
268
],
"spans": [
{
"bbox": [
216,
111,
395,
268
],
"score": 0.968,
"type": "image",
"image_path": "60b7d14ab4a629377a795bcc99485453a195556eaadd495831ef6798a0248ae6.jpg"
}
]
}
],
"index": 6.5,
"virtual_lines": [
{
"bbox": [
216,
111,
395,
124.08333333333333
],
"spans": [],
"index": 1
},
{
"bbox": [
216,
124.08333333333333,
395,
137.16666666666666
],
"spans": [],
"index": 2
},
{
"bbox": [
216,
137.16666666666666,
395,
150.25
],
"spans": [],
"index": 3
},
{
"bbox": [
216,
150.25,
395,
163.33333333333334
],
"spans": [],
"index": 4
},
{
"bbox": [
216,
163.33333333333334,
395,
176.41666666666669
],
"spans": [],
"index": 5
},
{
"bbox": [
216,
176.41666666666669,
395,
189.50000000000003
],
"spans": [],
"index": 6
},
{
"bbox": [
216,
189.50000000000003,
395,
202.58333333333337
],
"spans": [],
"index": 7
},
{
"bbox": [
216,
202.58333333333337,
395,
215.6666666666667
],
"spans": [],
"index": 8
},
{
"bbox": [
216,
215.6666666666667,
395,
228.75000000000006
],
"spans": [],
"index": 9
},
{
"bbox": [
216,
228.75000000000006,
395,
241.8333333333334
],
"spans": [],
"index": 10
},
{
"bbox": [
216,
241.8333333333334,
395,
254.91666666666674
],
"spans": [],
"index": 11
},
{
"bbox": [
216,
254.91666666666674,
395,
268.00000000000006
],
"spans": [],
"index": 12
}
]
},
{
"type": "image_caption",
"bbox": [
106,
286,
505,
354
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
287,
505,
300
],
"spans": [
{
"bbox": [
106,
287,
272,
300
],
"score": 1.0,
"content": "Figure A.1: Computation times for graph",
"type": "text"
},
{
"bbox": [
273,
288,
278,
298
],
"score": 0.61,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
278,
287,
505,
300
],
"score": 1.0,
"content": "-SNE and graph CNE with 2 and 128 output dimensions.",
"type": "text"
}
],
"index": 13
},
{
"bbox": [
105,
298,
505,
311
],
"spans": [
{
"bbox": [
105,
298,
242,
311
],
"score": 1.0,
"content": "openTSNE was run on CPU with",
"type": "text"
},
{
"bbox": [
243,
299,
294,
309
],
"score": 0.85,
"content": "\\mathrm { n - j } \\mathrm { o b s = - 1 }",
"type": "inline_equation"
},
{
"bbox": [
294,
298,
505,
311
],
"score": 1.0,
"content": ". CNE was run on GPU. Datasets are ordered by the",
"type": "text"
}
],
"index": 14
},
{
"bbox": [
105,
309,
505,
322
],
"spans": [
{
"bbox": [
105,
309,
505,
322
],
"score": 1.0,
"content": "increasing number of nodes. The runtime of openTSNE (for a given number of gradient descent",
"type": "text"
}
],
"index": 15
},
{
"bbox": [
105,
320,
506,
333
],
"spans": [
{
"bbox": [
105,
320,
506,
333
],
"score": 1.0,
"content": "steps) grows linearly with the number of nodes. The runtime of CNE (for a given number of epochs",
"type": "text"
}
],
"index": 16
},
{
"bbox": [
105,
331,
505,
344
],
"spans": [
{
"bbox": [
105,
331,
505,
344
],
"score": 1.0,
"content": "and a given batch size) grows linearly with the number of edges. The Pubmed dataset (PUB) has",
"type": "text"
}
],
"index": 17
},
{
"bbox": [
105,
342,
383,
354
],
"spans": [
{
"bbox": [
105,
342,
383,
354
],
"score": 1.0,
"content": "fewer edges than the Amazon datasets (APH and ACO), see Table 1.",
"type": "text"
}
],
"index": 18
}
],
"index": 15.5
}
],
"index": 11.0
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
174,
237,
446,
519
],
"blocks": [
{
"type": "image_caption",
"bbox": [
174,
230,
269,
239
],
"group_id": 0,
"lines": [
{
"bbox": [
174,
230,
270,
239
],
"spans": [
{
"bbox": [
174,
230,
270,
239
],
"score": 1.0,
"content": "ACO, per-node normalization",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image_caption",
"bbox": [
326,
231,
434,
239
],
"group_id": 0,
"lines": [
{
"bbox": [
326,
230,
434,
240
],
"spans": [
{
"bbox": [
326,
230,
434,
240
],
"score": 1.0,
"content": "ACO, whole-matrix normalization",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "image_body",
"bbox": [
174,
237,
446,
519
],
"group_id": 0,
"lines": [
{
"bbox": [
174,
237,
446,
519
],
"spans": [
{
"bbox": [
174,
237,
446,
519
],
"score": 0.796,
"type": "image",
"image_path": "3e059afd2ab2d936a50c61c839b21eda3094285bf8417f9e3a5dc2974b23345b.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
174,
237,
446,
249.26086956521738
],
"spans": [],
"index": 2
},
{
"bbox": [
174,
249.26086956521738,
446,
261.52173913043475
],
"spans": [],
"index": 3
},
{
"bbox": [
174,
261.52173913043475,
446,
273.78260869565213
],
"spans": [],
"index": 4
},
{
"bbox": [
174,
273.78260869565213,
446,
286.0434782608695
],
"spans": [],
"index": 5
},
{
"bbox": [
174,
286.0434782608695,
446,
298.3043478260869
],
"spans": [],
"index": 6
},
{
"bbox": [
174,
298.3043478260869,
446,
310.56521739130426
],
"spans": [],
"index": 7
},
{
"bbox": [
174,
310.56521739130426,
446,
322.82608695652164
],
"spans": [],
"index": 8
},
{
"bbox": [
174,
322.82608695652164,
446,
335.086956521739
],
"spans": [],
"index": 9
},
{
"bbox": [
174,
335.086956521739,
446,
347.3478260869564
],
"spans": [],
"index": 10
},
{
"bbox": [
174,
347.3478260869564,
446,
359.60869565217376
],
"spans": [],
"index": 11
},
{
"bbox": [
174,
359.60869565217376,
446,
371.86956521739114
],
"spans": [],
"index": 12
},
{
"bbox": [
174,
371.86956521739114,
446,
384.1304347826085
],
"spans": [],
"index": 13
},
{
"bbox": [
174,
384.1304347826085,
446,
396.3913043478259
],
"spans": [],
"index": 14
},
{
"bbox": [
174,
396.3913043478259,
446,
408.65217391304327
],
"spans": [],
"index": 15
},
{
"bbox": [
174,
408.65217391304327,
446,
420.91304347826065
],
"spans": [],
"index": 16
},
{
"bbox": [
174,
420.91304347826065,
446,
433.173913043478
],
"spans": [],
"index": 17
},
{
"bbox": [
174,
433.173913043478,
446,
445.4347826086954
],
"spans": [],
"index": 18
},
{
"bbox": [
174,
445.4347826086954,
446,
457.6956521739128
],
"spans": [],
"index": 19
},
{
"bbox": [
174,
457.6956521739128,
446,
469.95652173913015
],
"spans": [],
"index": 20
},
{
"bbox": [
174,
469.95652173913015,
446,
482.21739130434753
],
"spans": [],
"index": 21
},
{
"bbox": [
174,
482.21739130434753,
446,
494.4782608695649
],
"spans": [],
"index": 22
},
{
"bbox": [
174,
494.4782608695649,
446,
506.7391304347823
],
"spans": [],
"index": 23
},
{
"bbox": [
174,
506.7391304347823,
446,
518.9999999999997
],
"spans": [],
"index": 24
}
]
},
{
"type": "image_caption",
"bbox": [
106,
546,
505,
579
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
545,
505,
559
],
"spans": [
{
"bbox": [
105,
545,
185,
559
],
"score": 1.0,
"content": "Figure A.2: Graph",
"type": "text"
},
{
"bbox": [
186,
547,
191,
556
],
"score": 0.57,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
191,
545,
505,
559
],
"score": 1.0,
"content": "-SNE visualizations of ACO and APH datasets using per-node normalization",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
556,
505,
570
],
"spans": [
{
"bbox": [
105,
556,
505,
570
],
"score": 1.0,
"content": "of the adjacency matrix (default) and whole-matrix normalization. Embeddings in each row were",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
568,
242,
580
],
"spans": [
{
"bbox": [
106,
568,
242,
580
],
"score": 1.0,
"content": "aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 27
}
],
"index": 26
}
],
"index": 7.0
}
],
"page_idx": 13,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
313,
763
],
"spans": [
{
"bbox": [
299,
750,
313,
763
],
"score": 1.0,
"content": "14",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
174,
237,
446,
519
],
"blocks": [
{
"type": "image_caption",
"bbox": [
174,
230,
269,
239
],
"group_id": 0,
"lines": [
{
"bbox": [
174,
230,
270,
239
],
"spans": [
{
"bbox": [
174,
230,
270,
239
],
"score": 1.0,
"content": "ACO, per-node normalization",
"type": "text"
}
],
"index": 0
}
],
"index": 0
},
{
"type": "image_caption",
"bbox": [
326,
231,
434,
239
],
"group_id": 0,
"lines": [
{
"bbox": [
326,
230,
434,
240
],
"spans": [
{
"bbox": [
326,
230,
434,
240
],
"score": 1.0,
"content": "ACO, whole-matrix normalization",
"type": "text"
}
],
"index": 1
}
],
"index": 1
},
{
"type": "image_body",
"bbox": [
174,
237,
446,
519
],
"group_id": 0,
"lines": [
{
"bbox": [
174,
237,
446,
519
],
"spans": [
{
"bbox": [
174,
237,
446,
519
],
"score": 0.796,
"type": "image",
"image_path": "3e059afd2ab2d936a50c61c839b21eda3094285bf8417f9e3a5dc2974b23345b.jpg"
}
]
}
],
"index": 13,
"virtual_lines": [
{
"bbox": [
174,
237,
446,
249.26086956521738
],
"spans": [],
"index": 2
},
{
"bbox": [
174,
249.26086956521738,
446,
261.52173913043475
],
"spans": [],
"index": 3
},
{
"bbox": [
174,
261.52173913043475,
446,
273.78260869565213
],
"spans": [],
"index": 4
},
{
"bbox": [
174,
273.78260869565213,
446,
286.0434782608695
],
"spans": [],
"index": 5
},
{
"bbox": [
174,
286.0434782608695,
446,
298.3043478260869
],
"spans": [],
"index": 6
},
{
"bbox": [
174,
298.3043478260869,
446,
310.56521739130426
],
"spans": [],
"index": 7
},
{
"bbox": [
174,
310.56521739130426,
446,
322.82608695652164
],
"spans": [],
"index": 8
},
{
"bbox": [
174,
322.82608695652164,
446,
335.086956521739
],
"spans": [],
"index": 9
},
{
"bbox": [
174,
335.086956521739,
446,
347.3478260869564
],
"spans": [],
"index": 10
},
{
"bbox": [
174,
347.3478260869564,
446,
359.60869565217376
],
"spans": [],
"index": 11
},
{
"bbox": [
174,
359.60869565217376,
446,
371.86956521739114
],
"spans": [],
"index": 12
},
{
"bbox": [
174,
371.86956521739114,
446,
384.1304347826085
],
"spans": [],
"index": 13
},
{
"bbox": [
174,
384.1304347826085,
446,
396.3913043478259
],
"spans": [],
"index": 14
},
{
"bbox": [
174,
396.3913043478259,
446,
408.65217391304327
],
"spans": [],
"index": 15
},
{
"bbox": [
174,
408.65217391304327,
446,
420.91304347826065
],
"spans": [],
"index": 16
},
{
"bbox": [
174,
420.91304347826065,
446,
433.173913043478
],
"spans": [],
"index": 17
},
{
"bbox": [
174,
433.173913043478,
446,
445.4347826086954
],
"spans": [],
"index": 18
},
{
"bbox": [
174,
445.4347826086954,
446,
457.6956521739128
],
"spans": [],
"index": 19
},
{
"bbox": [
174,
457.6956521739128,
446,
469.95652173913015
],
"spans": [],
"index": 20
},
{
"bbox": [
174,
469.95652173913015,
446,
482.21739130434753
],
"spans": [],
"index": 21
},
{
"bbox": [
174,
482.21739130434753,
446,
494.4782608695649
],
"spans": [],
"index": 22
},
{
"bbox": [
174,
494.4782608695649,
446,
506.7391304347823
],
"spans": [],
"index": 23
},
{
"bbox": [
174,
506.7391304347823,
446,
518.9999999999997
],
"spans": [],
"index": 24
}
]
},
{
"type": "image_caption",
"bbox": [
106,
546,
505,
579
],
"group_id": 0,
"lines": [
{
"bbox": [
105,
545,
505,
559
],
"spans": [
{
"bbox": [
105,
545,
185,
559
],
"score": 1.0,
"content": "Figure A.2: Graph",
"type": "text"
},
{
"bbox": [
186,
547,
191,
556
],
"score": 0.57,
"content": "t { \\cdot }",
"type": "inline_equation"
},
{
"bbox": [
191,
545,
505,
559
],
"score": 1.0,
"content": "-SNE visualizations of ACO and APH datasets using per-node normalization",
"type": "text"
}
],
"index": 25
},
{
"bbox": [
105,
556,
505,
570
],
"spans": [
{
"bbox": [
105,
556,
505,
570
],
"score": 1.0,
"content": "of the adjacency matrix (default) and whole-matrix normalization. Embeddings in each row were",
"type": "text"
}
],
"index": 26
},
{
"bbox": [
106,
568,
242,
580
],
"spans": [
{
"bbox": [
106,
568,
242,
580
],
"score": 1.0,
"content": "aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 27
}
],
"index": 26
}
],
"index": 7.0
}
]
},
{
"preproc_blocks": [
{
"type": "image",
"bbox": [
106,
90,
489,
655
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
90,
489,
655
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
90,
489,
655
],
"spans": [
{
"bbox": [
106,
90,
489,
655
],
"score": 0.968,
"type": "image",
"image_path": "2e3a4e97efb0b5a71ebb8d946eb97878711824e2f09632ab76746aa0e846c712.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
106,
90,
489,
278.33333333333337
],
"spans": [],
"index": 0
},
{
"bbox": [
106,
278.33333333333337,
489,
466.66666666666674
],
"spans": [],
"index": 1
},
{
"bbox": [
106,
466.66666666666674,
489,
655.0000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
669,
505,
703
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
669,
504,
682
],
"spans": [
{
"bbox": [
106,
669,
504,
682
],
"score": 1.0,
"content": "Figure A.3: Embeddings of all considered datasets obtained using FDP (Fruchterman & Reingold,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
679,
505,
693
],
"spans": [
{
"bbox": [
106,
679,
271,
693
],
"score": 1.0,
"content": "1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
271,
681,
276,
691
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
276,
679,
440,
693
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), and our graph",
"type": "text"
},
{
"bbox": [
440,
681,
445,
690
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
446,
679,
505,
693
],
"score": 1.0,
"content": "-SNE. Embed-",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
691,
337,
704
],
"spans": [
{
"bbox": [
106,
691,
337,
704
],
"score": 1.0,
"content": "dings in each row were aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
}
],
"page_idx": 14,
"page_size": [
612,
792
],
"discarded_blocks": [
{
"type": "discarded",
"bbox": [
300,
751,
311,
760
],
"lines": [
{
"bbox": [
299,
750,
313,
764
],
"spans": [
{
"bbox": [
299,
750,
313,
764
],
"score": 1.0,
"content": "15",
"type": "text"
}
]
}
]
},
{
"type": "discarded",
"bbox": [
106,
27,
308,
37
],
"lines": [
{
"bbox": [
106,
26,
308,
38
],
"spans": [
{
"bbox": [
106,
26,
308,
38
],
"score": 1.0,
"content": "Under review as a conference paper at ICLR 2024",
"type": "text"
}
]
}
]
}
],
"para_blocks": [
{
"type": "image",
"bbox": [
106,
90,
489,
655
],
"blocks": [
{
"type": "image_body",
"bbox": [
106,
90,
489,
655
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
90,
489,
655
],
"spans": [
{
"bbox": [
106,
90,
489,
655
],
"score": 0.968,
"type": "image",
"image_path": "2e3a4e97efb0b5a71ebb8d946eb97878711824e2f09632ab76746aa0e846c712.jpg"
}
]
}
],
"index": 1,
"virtual_lines": [
{
"bbox": [
106,
90,
489,
278.33333333333337
],
"spans": [],
"index": 0
},
{
"bbox": [
106,
278.33333333333337,
489,
466.66666666666674
],
"spans": [],
"index": 1
},
{
"bbox": [
106,
466.66666666666674,
489,
655.0000000000001
],
"spans": [],
"index": 2
}
]
},
{
"type": "image_caption",
"bbox": [
106,
669,
505,
703
],
"group_id": 0,
"lines": [
{
"bbox": [
106,
669,
504,
682
],
"spans": [
{
"bbox": [
106,
669,
504,
682
],
"score": 1.0,
"content": "Figure A.3: Embeddings of all considered datasets obtained using FDP (Fruchterman & Reingold,",
"type": "text"
}
],
"index": 3
},
{
"bbox": [
106,
679,
505,
693
],
"spans": [
{
"bbox": [
106,
679,
271,
693
],
"score": 1.0,
"content": "1991), DRGraph (Zhu et al., 2020a), and",
"type": "text"
},
{
"bbox": [
271,
681,
276,
691
],
"score": 0.76,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
276,
679,
440,
693
],
"score": 1.0,
"content": "-FDP (Zhong et al., 2023), and our graph",
"type": "text"
},
{
"bbox": [
440,
681,
445,
690
],
"score": 0.52,
"content": "t",
"type": "inline_equation"
},
{
"bbox": [
446,
679,
505,
693
],
"score": 1.0,
"content": "-SNE. Embed-",
"type": "text"
}
],
"index": 4
},
{
"bbox": [
106,
691,
337,
704
],
"spans": [
{
"bbox": [
106,
691,
337,
704
],
"score": 1.0,
"content": "dings in each row were aligned using Procrustes rotation.",
"type": "text"
}
],
"index": 5
}
],
"index": 4
}
],
"index": 2.5
}
]
}
],
"_backend": "pipeline",
"_version_name": "2.1.11"
}