diff --git "a/parse/train/SkNQeiRpb/SkNQeiRpb_middle.json" "b/parse/train/SkNQeiRpb/SkNQeiRpb_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/SkNQeiRpb/SkNQeiRpb_middle.json"
@@ -0,0 +1,21442 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 78,
+ 503,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 505,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 505,
+ 98
+ ],
+ "score": 1.0,
+ "content": "TRAINING DEEP AUTOENCODERS FOR RECOM-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 99,
+ 246,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 99,
+ 246,
+ 117
+ ],
+ "score": 1.0,
+ "content": "MENDER SYSTEMS",
+ "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,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 217,
+ 468,
+ 338
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 217,
+ 470,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 217,
+ 470,
+ 230
+ ],
+ "score": 1.0,
+ "content": "This paper proposes a new model for the rating prediction task in recommender",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 228,
+ 470,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 228,
+ 470,
+ 241
+ ],
+ "score": 1.0,
+ "content": "systems which significantly outperforms previous state-of-the art models on a",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 239,
+ 470,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 239,
+ 470,
+ 252
+ ],
+ "score": 1.0,
+ "content": "time-split Netflix data set. Our model is based on deep autoencoder with 6 lay-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 251,
+ 469,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 251,
+ 469,
+ 262
+ ],
+ "score": 1.0,
+ "content": "ers and is trained end-to-end without any layer-wise pre-training. We empirically",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 142,
+ 261,
+ 469,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 261,
+ 469,
+ 273
+ ],
+ "score": 1.0,
+ "content": "demonstrate that: a) deep autoencoder models generalize much better than the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 272,
+ 470,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 272,
+ 470,
+ 284
+ ],
+ "score": 1.0,
+ "content": "shallow ones, b) non-linear activation functions with negative parts are crucial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 283,
+ 470,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 283,
+ 470,
+ 296
+ ],
+ "score": 1.0,
+ "content": "for training deep models, and c) heavy use of regularization techniques such as",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 294,
+ 470,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 294,
+ 470,
+ 307
+ ],
+ "score": 1.0,
+ "content": "dropout is necessary to prevent overfitting. We also propose a new training al-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 306,
+ 470,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 306,
+ 470,
+ 317
+ ],
+ "score": 1.0,
+ "content": "gorithm based on iterative output re-feeding to overcome natural sparseness of",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "score": 1.0,
+ "content": "collaborate filtering. The new algorithm significantly speeds up training and im-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 328,
+ 256,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 328,
+ 256,
+ 339
+ ],
+ "score": 1.0,
+ "content": "proves model performance.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 374,
+ 206,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 208,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 208,
+ 390
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 405,
+ 504,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 404,
+ 504,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 404,
+ 504,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Sites like Amazon, Netflix and Spotify use recommender systems to suggest items to users. Recom-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "score": 1.0,
+ "content": "mender systems can be divided into two categories: context-based and personalized recommenda-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 132,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 132,
+ 439
+ ],
+ "score": 1.0,
+ "content": "tions.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 505,
+ 509
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Context based recommendations take into account contextual factors such as location, date and time",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(Adomavicius & Tuzhilin, 2011). Personalized recommendations typically suggest items to users",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "using the collaborative filtering (CF) approach. In this approach the user’s interests are predicted",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "based on the analysis of tastes and preference of other users in the system and implicitly inferring",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "“similarity” between them. The underlying assumption is that two people who have similar tastes,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 500,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 500,
+ 511
+ ],
+ "score": 1.0,
+ "content": "have a higher likelihood of having the same opinion on an item than two randomly chosen people.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 504,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "In designing recommender systems, the goal is to improve the accuracy of predictions. The Netflix",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "Prize contest provides the most famous example of this problem (Bennett et al., 2007): Netflix held",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "the Netflix Prize to substantially improve the accuracy of the algorithm to predict user ratings for",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 504,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 439,
+ 562
+ ],
+ "score": 1.0,
+ "content": "films. This is a classic CF problem: Infer the missing entries in an mxn matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 549,
+ 449,
+ 559
+ ],
+ "score": 0.69,
+ "content": "R",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 547,
+ 482,
+ 562
+ ],
+ "score": 1.0,
+ "content": ", whose",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 548,
+ 504,
+ 560
+ ],
+ "score": 0.91,
+ "content": "( i , j )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 324,
+ 572
+ ],
+ "score": 1.0,
+ "content": "entry describes the ratings given by the ith user to the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 560,
+ 330,
+ 571
+ ],
+ "score": 0.53,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 559,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "th item. The performance is then measured",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 273,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 273,
+ 582
+ ],
+ "score": 1.0,
+ "content": "using Root Mean Squared Error (RMSE).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Training very deep autoencoders is non trivial both from optimization and regularization points of",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "view. Early works on training auto-enocoders adapted layer-wise pre-training to solve optimiza-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "tion issues (Hinton & Salakhutdinov, 2006). In this work, we empirically show that optimization",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 619,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 619,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "difficulties of training deep autoencoders can be solved by using scaled exponential linear units",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 107,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 631,
+ 140,
+ 642
+ ],
+ "score": 0.39,
+ "content": "( S E L U s ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": ")(Klambauer et al., 2017). This enables training without any layer-wise pre-training or resid-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "ual connections. Since publicly available data sets for CF are relatively small, sufficiently large",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "models can easily overfit. To prevent overfitting we employ heavy dropout with drop probability as",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 663,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "high as 0.8. We also introduce a new output re-feeding training algorithm which helps to bypass",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 674,
+ 506,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 674,
+ 506,
+ 687
+ ],
+ "score": 1.0,
+ "content": "the natural sparseness of updates in collaborative filtering and helps to further improve the model",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 162,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 162,
+ 698
+ ],
+ "score": 1.0,
+ "content": "performance.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 119,
+ 722,
+ 361,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 362,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 362,
+ 733
+ ],
+ "score": 1.0,
+ "content": "1Our code is available at https://github.com/Anonymous",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 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": [
+ 108,
+ 78,
+ 503,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 505,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 78,
+ 505,
+ 98
+ ],
+ "score": 1.0,
+ "content": "TRAINING DEEP AUTOENCODERS FOR RECOM-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 99,
+ 246,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 99,
+ 246,
+ 117
+ ],
+ "score": 1.0,
+ "content": "MENDER SYSTEMS",
+ "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,
+ 187,
+ 333,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 186,
+ 335,
+ 200
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 217,
+ 468,
+ 338
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 142,
+ 217,
+ 470,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 217,
+ 470,
+ 230
+ ],
+ "score": 1.0,
+ "content": "This paper proposes a new model for the rating prediction task in recommender",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 141,
+ 228,
+ 470,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 228,
+ 470,
+ 241
+ ],
+ "score": 1.0,
+ "content": "systems which significantly outperforms previous state-of-the art models on a",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 239,
+ 470,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 239,
+ 470,
+ 252
+ ],
+ "score": 1.0,
+ "content": "time-split Netflix data set. Our model is based on deep autoencoder with 6 lay-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 141,
+ 251,
+ 469,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 251,
+ 469,
+ 262
+ ],
+ "score": 1.0,
+ "content": "ers and is trained end-to-end without any layer-wise pre-training. We empirically",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 142,
+ 261,
+ 469,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 261,
+ 469,
+ 273
+ ],
+ "score": 1.0,
+ "content": "demonstrate that: a) deep autoencoder models generalize much better than the",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 272,
+ 470,
+ 284
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 272,
+ 470,
+ 284
+ ],
+ "score": 1.0,
+ "content": "shallow ones, b) non-linear activation functions with negative parts are crucial",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 283,
+ 470,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 283,
+ 470,
+ 296
+ ],
+ "score": 1.0,
+ "content": "for training deep models, and c) heavy use of regularization techniques such as",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 141,
+ 294,
+ 470,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 294,
+ 470,
+ 307
+ ],
+ "score": 1.0,
+ "content": "dropout is necessary to prevent overfitting. We also propose a new training al-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 306,
+ 470,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 306,
+ 470,
+ 317
+ ],
+ "score": 1.0,
+ "content": "gorithm based on iterative output re-feeding to overcome natural sparseness of",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 316,
+ 469,
+ 329
+ ],
+ "score": 1.0,
+ "content": "collaborate filtering. The new algorithm significantly speeds up training and im-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 328,
+ 256,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 328,
+ 256,
+ 339
+ ],
+ "score": 1.0,
+ "content": "proves model performance.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 141,
+ 217,
+ 470,
+ 339
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 374,
+ 206,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 208,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 373,
+ 208,
+ 390
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 405,
+ 504,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 404,
+ 504,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 404,
+ 504,
+ 417
+ ],
+ "score": 1.0,
+ "content": "Sites like Amazon, Netflix and Spotify use recommender systems to suggest items to users. Recom-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 416,
+ 504,
+ 428
+ ],
+ "score": 1.0,
+ "content": "mender systems can be divided into two categories: context-based and personalized recommenda-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 426,
+ 132,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 426,
+ 132,
+ 439
+ ],
+ "score": 1.0,
+ "content": "tions.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 404,
+ 504,
+ 439
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 443,
+ 505,
+ 509
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 505,
+ 456
+ ],
+ "score": 1.0,
+ "content": "Context based recommendations take into account contextual factors such as location, date and time",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 454,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 454,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "(Adomavicius & Tuzhilin, 2011). Personalized recommendations typically suggest items to users",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 505,
+ 478
+ ],
+ "score": 1.0,
+ "content": "using the collaborative filtering (CF) approach. In this approach the user’s interests are predicted",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 505,
+ 491
+ ],
+ "score": 1.0,
+ "content": "based on the analysis of tastes and preference of other users in the system and implicitly inferring",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 487,
+ 505,
+ 500
+ ],
+ "score": 1.0,
+ "content": "“similarity” between them. The underlying assumption is that two people who have similar tastes,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 498,
+ 500,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 498,
+ 500,
+ 511
+ ],
+ "score": 1.0,
+ "content": "have a higher likelihood of having the same opinion on an item than two randomly chosen people.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 443,
+ 505,
+ 511
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 515,
+ 504,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "In designing recommender systems, the goal is to improve the accuracy of predictions. The Netflix",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "Prize contest provides the most famous example of this problem (Bennett et al., 2007): Netflix held",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 537,
+ 506,
+ 550
+ ],
+ "score": 1.0,
+ "content": "the Netflix Prize to substantially improve the accuracy of the algorithm to predict user ratings for",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 547,
+ 504,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 547,
+ 439,
+ 562
+ ],
+ "score": 1.0,
+ "content": "films. This is a classic CF problem: Infer the missing entries in an mxn matrix,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 440,
+ 549,
+ 449,
+ 559
+ ],
+ "score": 0.69,
+ "content": "R",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 547,
+ 482,
+ 562
+ ],
+ "score": 1.0,
+ "content": ", whose",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 548,
+ 504,
+ 560
+ ],
+ "score": 0.91,
+ "content": "( i , j )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 506,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 324,
+ 572
+ ],
+ "score": 1.0,
+ "content": "entry describes the ratings given by the ith user to the",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 325,
+ 560,
+ 330,
+ 571
+ ],
+ "score": 0.53,
+ "content": "j",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 331,
+ 559,
+ 506,
+ 572
+ ],
+ "score": 1.0,
+ "content": "th item. The performance is then measured",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 273,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 273,
+ 582
+ ],
+ "score": 1.0,
+ "content": "using Root Mean Squared Error (RMSE).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 515,
+ 506,
+ 582
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 697
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Training very deep autoencoders is non trivial both from optimization and regularization points of",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 598,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "view. Early works on training auto-enocoders adapted layer-wise pre-training to solve optimiza-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "tion issues (Hinton & Salakhutdinov, 2006). In this work, we empirically show that optimization",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 104,
+ 619,
+ 505,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 619,
+ 505,
+ 633
+ ],
+ "score": 1.0,
+ "content": "difficulties of training deep autoencoders can be solved by using scaled exponential linear units",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 107,
+ 631,
+ 505,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 631,
+ 140,
+ 642
+ ],
+ "score": 0.39,
+ "content": "( S E L U s ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 631,
+ 505,
+ 644
+ ],
+ "score": 1.0,
+ "content": ")(Klambauer et al., 2017). This enables training without any layer-wise pre-training or resid-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 641,
+ 505,
+ 656
+ ],
+ "score": 1.0,
+ "content": "ual connections. Since publicly available data sets for CF are relatively small, sufficiently large",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 653,
+ 505,
+ 666
+ ],
+ "score": 1.0,
+ "content": "models can easily overfit. To prevent overfitting we employ heavy dropout with drop probability as",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 663,
+ 506,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 506,
+ 677
+ ],
+ "score": 1.0,
+ "content": "high as 0.8. We also introduce a new output re-feeding training algorithm which helps to bypass",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 674,
+ 506,
+ 687
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 674,
+ 506,
+ 687
+ ],
+ "score": 1.0,
+ "content": "the natural sparseness of updates in collaborative filtering and helps to further improve the model",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 162,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 162,
+ 698
+ ],
+ "score": 1.0,
+ "content": "performance.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 104,
+ 588,
+ 506,
+ 698
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 200,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 203,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 203,
+ 95
+ ],
+ "score": 1.0,
+ "content": "1.1 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 111,
+ 505,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Deep learning (LeCun et al., 2015) has led to breakthroughs in image recognition, natural language",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "understanding, and reinforcement learning. Naturally, these successes fuel an interest for using",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 133,
+ 505,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 133,
+ 505,
+ 146
+ ],
+ "score": 1.0,
+ "content": "deep learning in recommender systems. First attempts at using deep learning for recommender",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "systems involved restricted Boltzman machines (RBM) (Salakhutdinov et al., 2007). Several re-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 506,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 506,
+ 168
+ ],
+ "score": 1.0,
+ "content": "cent approaches use autoencoders (Sedhain et al., 2015; Strub & Mary, 2015), feed-forward neural",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 506,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 506,
+ 179
+ ],
+ "score": 1.0,
+ "content": "networks (He et al., 2017), neural autoregressive architectures (Zheng et al., 2016) and recurrent",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "recommender networks (Wu et al., 2017). Many popular matrix factorization techniques can be",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "thought of as a form of dimensionality reduction. It is, therefore, natural to adapt deep autoencoders",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 356,
+ 212
+ ],
+ "score": 1.0,
+ "content": "for this task as well. I-AutoRec (item-based autoencoder) and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 200,
+ 364,
+ 209
+ ],
+ "score": 0.41,
+ "content": "U _ { ☉ }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "-AutoRec (user-based autoencoder)",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "are first successful attempts to do so Sedhain et al. (2015). Stacked de-noising autoencoders has",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 222,
+ 411,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 411,
+ 233
+ ],
+ "score": 1.0,
+ "content": "been sucesfully used on this task as well (Li et al., 2015; Wang et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 237,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 504,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 504,
+ 250
+ ],
+ "score": 1.0,
+ "content": "There are many non deep learning types of approaches to collaborative filtering (CF) (Breese et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "1998; Ricci et al., 2011). Matrix factorization techniques, such as alternating least squares (ALS)",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "(Kim & Park, 2008; Koren et al., 2009) and probabilistic matrix factorization (Mnih & Salakhutdi-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "nov, 2008) are particularly popular. The most robust systems may incorporate several ideas together",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "such as the winning solution to the Netflix Prize competition (Koren, 2009). Note that Netflix Prize",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "data also includes temporal signal - time when each rating has been made. Thus, several classic",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 465,
+ 316
+ ],
+ "score": 1.0,
+ "content": "CF approaches has been extended to incorporate temporal information such as TimeSVD",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 304,
+ 477,
+ 314
+ ],
+ "score": 0.55,
+ "content": "^ { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Koren",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "(2010), as well as more recent RNN-based techniques such as recurrent recommender networks Wu",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 159,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 159,
+ 337
+ ],
+ "score": 1.0,
+ "content": "et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 374,
+ 164,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 167,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 167,
+ 389
+ ],
+ "score": 1.0,
+ "content": "2 MODEL",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 411,
+ 505,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 424
+ ],
+ "score": 1.0,
+ "content": "Our model is inspired by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 412,
+ 219,
+ 421
+ ],
+ "score": 0.47,
+ "content": "U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 411,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "-AutoRec approach with several important distinctions. We train much",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "deeper models. To enable this without any pre-training, we: a) use “scaled exponential linear units”",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "(SELUs) Klambauer et al. (2017), b) use high dropout rates, and d) use iterative output re-feeding",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 172,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 172,
+ 459
+ ],
+ "score": 1.0,
+ "content": "during training.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 461,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 424,
+ 474
+ ],
+ "score": 1.0,
+ "content": "An autoencoder is a network which implements two transformations - encoder",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 0.86,
+ "content": "e n c o d e ( x ) : R ^ { n } \\to",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 107,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 472,
+ 120,
+ 483
+ ],
+ "score": 0.87,
+ "content": "R ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 120,
+ 471,
+ 138,
+ 486
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 472,
+ 234,
+ 484
+ ],
+ "score": 0.81,
+ "content": "d e c o d e r ( z ) : R ^ { d } \\to R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 471,
+ 388,
+ 486
+ ],
+ "score": 1.0,
+ "content": ". The “goal” of autoenoder is to obtain",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 473,
+ 394,
+ 482
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 395,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "dimensional representation",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 280,
+ 496
+ ],
+ "score": 1.0,
+ "content": "of data such that an error measure between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 485,
+ 287,
+ 493
+ ],
+ "score": 0.76,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 484,
+ 305,
+ 496
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 483,
+ 419,
+ 496
+ ],
+ "score": 0.9,
+ "content": "f ( x ) = d e c o d e ( e n c o d e ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 484,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "is minimized Hinton",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "& Zemel (1994). Figure 1 depicts typical 4-layer autoencoder network. If noise is added to the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 519
+ ],
+ "score": 1.0,
+ "content": "data during encoding step, the autoencoder is called de-noising. Autoencoder is an excellent tool",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "for dimensionality reduction and can be thought of as a strict generalization of principle component",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "analysis (PCA) Hinton & Salakhutdinov (2006). An autoencoder without non-linear activations",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "and only with “code” layer should be able to learn PCA transformation in the encoder if trained to",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 272,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 272,
+ 561
+ ],
+ "score": 1.0,
+ "content": "optimize mean squared error (MSE) loss.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 566,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "In our model, both encoder and decoder parts of the autoencoder consist of feed-forward neural",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 338,
+ 589
+ ],
+ "score": 1.0,
+ "content": "networks with classical fully connected layers computing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 577,
+ 412,
+ 589
+ ],
+ "score": 0.92,
+ "content": "l = f ( W * x + b )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 577,
+ 443,
+ 589
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 578,
+ 450,
+ 588
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "is some non-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "linear activation function. If range of the activation function is smaller than that of data, the last layer",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 486,
+ 612
+ ],
+ "score": 1.0,
+ "content": "of the decoder should be kept linear. We found it to be very important for activation function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 599,
+ 493,
+ 611
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "in",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "hidden layers to contain non-zero negative part, and we use SELU units in most of our experiments",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 220,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 220,
+ 633
+ ],
+ "score": 1.0,
+ "content": "(see Section 3.2 for details).",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 638,
+ 504,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 650
+ ],
+ "score": 1.0,
+ "content": "If decoder mirrors encoder architecture (as it does in our model), then one can constrain decoder’s",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 140,
+ 661
+ ],
+ "score": 1.0,
+ "content": "weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 648,
+ 156,
+ 661
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { W } _ { d } ^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 648,
+ 333,
+ 661
+ ],
+ "score": 1.0,
+ "content": "to be equal to transposed encoder weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 648,
+ 348,
+ 661
+ ],
+ "score": 0.91,
+ "content": "W _ { e } ^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 348,
+ 648,
+ 471,
+ 661
+ ],
+ "score": 1.0,
+ "content": "from the corresponding layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 649,
+ 476,
+ 659
+ ],
+ "score": 0.49,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 648,
+ 506,
+ 661
+ ],
+ "score": 1.0,
+ "content": ". Such",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "autoencoder is called constrained or tied and has almost two times less free parameters than uncon-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 160,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 160,
+ 682
+ ],
+ "score": 1.0,
+ "content": "strained one.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Forward pass and inference. During forward pass (and inference) the model takes user represented",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 289,
+ 711
+ ],
+ "score": 1.0,
+ "content": "by his vector of ratings from the training set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 699,
+ 322,
+ 709
+ ],
+ "score": 0.89,
+ "content": "x \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 698,
+ 354,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 701,
+ 362,
+ 709
+ ],
+ "score": 0.76,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 698,
+ 487,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is number of items. Note that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 700,
+ 494,
+ 709
+ ],
+ "score": 0.76,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 285,
+ 721
+ ],
+ "score": 1.0,
+ "content": "very sparse, while the output of the decoder,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 710,
+ 330,
+ 722
+ ],
+ "score": 0.92,
+ "content": "f ( x ) \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 710,
+ 506,
+ 721
+ ],
+ "score": 1.0,
+ "content": "is dense and contains rating predictions for",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 200,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 200,
+ 733
+ ],
+ "score": 1.0,
+ "content": "all items in the corpus.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 200,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 203,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 82,
+ 203,
+ 95
+ ],
+ "score": 1.0,
+ "content": "1.1 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 111,
+ 505,
+ 232
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 110,
+ 505,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Deep learning (LeCun et al., 2015) has led to breakthroughs in image recognition, natural language",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 122,
+ 506,
+ 136
+ ],
+ "score": 1.0,
+ "content": "understanding, and reinforcement learning. Naturally, these successes fuel an interest for using",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 133,
+ 505,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 133,
+ 505,
+ 146
+ ],
+ "score": 1.0,
+ "content": "deep learning in recommender systems. First attempts at using deep learning for recommender",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "systems involved restricted Boltzman machines (RBM) (Salakhutdinov et al., 2007). Several re-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 506,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 506,
+ 168
+ ],
+ "score": 1.0,
+ "content": "cent approaches use autoencoders (Sedhain et al., 2015; Strub & Mary, 2015), feed-forward neural",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 167,
+ 506,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 167,
+ 506,
+ 179
+ ],
+ "score": 1.0,
+ "content": "networks (He et al., 2017), neural autoregressive architectures (Zheng et al., 2016) and recurrent",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 505,
+ 190
+ ],
+ "score": 1.0,
+ "content": "recommender networks (Wu et al., 2017). Many popular matrix factorization techniques can be",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 505,
+ 201
+ ],
+ "score": 1.0,
+ "content": "thought of as a form of dimensionality reduction. It is, therefore, natural to adapt deep autoencoders",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 199,
+ 505,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 199,
+ 356,
+ 212
+ ],
+ "score": 1.0,
+ "content": "for this task as well. I-AutoRec (item-based autoencoder) and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 356,
+ 200,
+ 364,
+ 209
+ ],
+ "score": 0.41,
+ "content": "U _ { ☉ }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 199,
+ 505,
+ 212
+ ],
+ "score": 1.0,
+ "content": "-AutoRec (user-based autoencoder)",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 505,
+ 222
+ ],
+ "score": 1.0,
+ "content": "are first successful attempts to do so Sedhain et al. (2015). Stacked de-noising autoencoders has",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 222,
+ 411,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 222,
+ 411,
+ 233
+ ],
+ "score": 1.0,
+ "content": "been sucesfully used on this task as well (Li et al., 2015; Wang et al., 2015).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 110,
+ 506,
+ 233
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 237,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 504,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 238,
+ 504,
+ 250
+ ],
+ "score": 1.0,
+ "content": "There are many non deep learning types of approaches to collaborative filtering (CF) (Breese et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 248,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "1998; Ricci et al., 2011). Matrix factorization techniques, such as alternating least squares (ALS)",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 259,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "(Kim & Park, 2008; Koren et al., 2009) and probabilistic matrix factorization (Mnih & Salakhutdi-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 271,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "nov, 2008) are particularly popular. The most robust systems may incorporate several ideas together",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 282,
+ 506,
+ 295
+ ],
+ "score": 1.0,
+ "content": "such as the winning solution to the Netflix Prize competition (Koren, 2009). Note that Netflix Prize",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 292,
+ 505,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 292,
+ 505,
+ 306
+ ],
+ "score": 1.0,
+ "content": "data also includes temporal signal - time when each rating has been made. Thus, several classic",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 303,
+ 506,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 303,
+ 465,
+ 316
+ ],
+ "score": 1.0,
+ "content": "CF approaches has been extended to incorporate temporal information such as TimeSVD",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 304,
+ 477,
+ 314
+ ],
+ "score": 0.55,
+ "content": "^ { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 478,
+ 303,
+ 506,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Koren",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 506,
+ 327
+ ],
+ "score": 1.0,
+ "content": "(2010), as well as more recent RNN-based techniques such as recurrent recommender networks Wu",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 159,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 159,
+ 337
+ ],
+ "score": 1.0,
+ "content": "et al. (2017).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 105,
+ 238,
+ 506,
+ 337
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 374,
+ 164,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 167,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 371,
+ 167,
+ 389
+ ],
+ "score": 1.0,
+ "content": "2 MODEL",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 411,
+ 505,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 505,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 210,
+ 424
+ ],
+ "score": 1.0,
+ "content": "Our model is inspired by",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 211,
+ 412,
+ 219,
+ 421
+ ],
+ "score": 0.47,
+ "content": "U",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 411,
+ 505,
+ 424
+ ],
+ "score": 1.0,
+ "content": "-AutoRec approach with several important distinctions. We train much",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 505,
+ 435
+ ],
+ "score": 1.0,
+ "content": "deeper models. To enable this without any pre-training, we: a) use “scaled exponential linear units”",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "(SELUs) Klambauer et al. (2017), b) use high dropout rates, and d) use iterative output re-feeding",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 444,
+ 172,
+ 459
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 444,
+ 172,
+ 459
+ ],
+ "score": 1.0,
+ "content": "during training.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 105,
+ 411,
+ 505,
+ 459
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 461,
+ 505,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 505,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 424,
+ 474
+ ],
+ "score": 1.0,
+ "content": "An autoencoder is a network which implements two transformations - encoder",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 424,
+ 461,
+ 505,
+ 474
+ ],
+ "score": 0.86,
+ "content": "e n c o d e ( x ) : R ^ { n } \\to",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 107,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 472,
+ 120,
+ 483
+ ],
+ "score": 0.87,
+ "content": "R ^ { d }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 120,
+ 471,
+ 138,
+ 486
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 138,
+ 472,
+ 234,
+ 484
+ ],
+ "score": 0.81,
+ "content": "d e c o d e r ( z ) : R ^ { d } \\to R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 235,
+ 471,
+ 388,
+ 486
+ ],
+ "score": 1.0,
+ "content": ". The “goal” of autoenoder is to obtain",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 473,
+ 394,
+ 482
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 395,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "dimensional representation",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 505,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 484,
+ 280,
+ 496
+ ],
+ "score": 1.0,
+ "content": "of data such that an error measure between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 485,
+ 287,
+ 493
+ ],
+ "score": 0.76,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 288,
+ 484,
+ 305,
+ 496
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 483,
+ 419,
+ 496
+ ],
+ "score": 0.9,
+ "content": "f ( x ) = d e c o d e ( e n c o d e ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 419,
+ 484,
+ 505,
+ 496
+ ],
+ "score": 1.0,
+ "content": "is minimized Hinton",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 506,
+ 508
+ ],
+ "score": 1.0,
+ "content": "& Zemel (1994). Figure 1 depicts typical 4-layer autoencoder network. If noise is added to the",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 519
+ ],
+ "score": 1.0,
+ "content": "data during encoding step, the autoencoder is called de-noising. Autoencoder is an excellent tool",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "for dimensionality reduction and can be thought of as a strict generalization of principle component",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 506,
+ 540
+ ],
+ "score": 1.0,
+ "content": "analysis (PCA) Hinton & Salakhutdinov (2006). An autoencoder without non-linear activations",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 538,
+ 505,
+ 550
+ ],
+ "score": 1.0,
+ "content": "and only with “code” layer should be able to learn PCA transformation in the encoder if trained to",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 272,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 272,
+ 561
+ ],
+ "score": 1.0,
+ "content": "optimize mean squared error (MSE) loss.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 461,
+ 506,
+ 561
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 566,
+ 505,
+ 632
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 565,
+ 505,
+ 577
+ ],
+ "score": 1.0,
+ "content": "In our model, both encoder and decoder parts of the autoencoder consist of feed-forward neural",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 505,
+ 589
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 338,
+ 589
+ ],
+ "score": 1.0,
+ "content": "networks with classical fully connected layers computing",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 577,
+ 412,
+ 589
+ ],
+ "score": 0.92,
+ "content": "l = f ( W * x + b )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 577,
+ 443,
+ 589
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 443,
+ 578,
+ 450,
+ 588
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 451,
+ 577,
+ 505,
+ 589
+ ],
+ "score": 1.0,
+ "content": "is some non-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 588,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "linear activation function. If range of the activation function is smaller than that of data, the last layer",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 505,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 599,
+ 486,
+ 612
+ ],
+ "score": 1.0,
+ "content": "of the decoder should be kept linear. We found it to be very important for activation function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 486,
+ 599,
+ 493,
+ 611
+ ],
+ "score": 0.84,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 494,
+ 599,
+ 505,
+ 612
+ ],
+ "score": 1.0,
+ "content": "in",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 609,
+ 505,
+ 623
+ ],
+ "score": 1.0,
+ "content": "hidden layers to contain non-zero negative part, and we use SELU units in most of our experiments",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 220,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 621,
+ 220,
+ 633
+ ],
+ "score": 1.0,
+ "content": "(see Section 3.2 for details).",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 565,
+ 505,
+ 633
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 638,
+ 504,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 506,
+ 650
+ ],
+ "score": 1.0,
+ "content": "If decoder mirrors encoder architecture (as it does in our model), then one can constrain decoder’s",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 506,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 140,
+ 661
+ ],
+ "score": 1.0,
+ "content": "weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 141,
+ 648,
+ 156,
+ 661
+ ],
+ "score": 0.91,
+ "content": "\\boldsymbol { W } _ { d } ^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 648,
+ 333,
+ 661
+ ],
+ "score": 1.0,
+ "content": "to be equal to transposed encoder weights",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 333,
+ 648,
+ 348,
+ 661
+ ],
+ "score": 0.91,
+ "content": "W _ { e } ^ { l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 348,
+ 648,
+ 471,
+ 661
+ ],
+ "score": 1.0,
+ "content": "from the corresponding layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 471,
+ 649,
+ 476,
+ 659
+ ],
+ "score": 0.49,
+ "content": "l",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 476,
+ 648,
+ 506,
+ 661
+ ],
+ "score": 1.0,
+ "content": ". Such",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 506,
+ 672
+ ],
+ "score": 1.0,
+ "content": "autoencoder is called constrained or tied and has almost two times less free parameters than uncon-",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 671,
+ 160,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 160,
+ 682
+ ],
+ "score": 1.0,
+ "content": "strained one.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 637,
+ 506,
+ 682
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Forward pass and inference. During forward pass (and inference) the model takes user represented",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 289,
+ 711
+ ],
+ "score": 1.0,
+ "content": "by his vector of ratings from the training set",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 289,
+ 699,
+ 322,
+ 709
+ ],
+ "score": 0.89,
+ "content": "x \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 323,
+ 698,
+ 354,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 355,
+ 701,
+ 362,
+ 709
+ ],
+ "score": 0.76,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 362,
+ 698,
+ 487,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is number of items. Note that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 487,
+ 700,
+ 494,
+ 709
+ ],
+ "score": 0.76,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 495,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 285,
+ 721
+ ],
+ "score": 1.0,
+ "content": "very sparse, while the output of the decoder,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 285,
+ 710,
+ 330,
+ 722
+ ],
+ "score": 0.92,
+ "content": "f ( x ) \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 330,
+ 710,
+ 506,
+ 721
+ ],
+ "score": 1.0,
+ "content": "is dense and contains rating predictions for",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 200,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 200,
+ 733
+ ],
+ "score": 1.0,
+ "content": "all items in the corpus.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 46.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 82,
+ 201,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 202,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 202,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2.1 LOSS FUNCTION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 103,
+ 504,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 400,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Since it doesn’t make sense to predict zeros in user’s representation vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 400,
+ 106,
+ 407,
+ 113
+ ],
+ "score": 0.78,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 103,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": ", we follow the approach",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 113,
+ 407,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 113,
+ 407,
+ 127
+ ],
+ "score": 1.0,
+ "content": "from Sedhain et al. (2015) and optimize Masked Mean Squared Error loss:",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "score": 0.95,
+ "content": "M M S E = \\frac { m _ { i } * ( r _ { i } - y _ { i } ) ^ { 2 } } { \\sum _ { i = 0 } ^ { i = n } m _ { i } }",
+ "type": "interline_equation",
+ "image_path": "1897d4b1db1749e191b172c3dd7d98ae749d38dd07fd68a752082ec3404f525a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 145.5
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 247,
+ 145.5,
+ 364,
+ 161.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 164,
+ 504,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 133,
+ 177
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 142,
+ 176
+ ],
+ "score": 0.86,
+ "content": "r _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 164,
+ 206,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is actual rating,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 206,
+ 167,
+ 215,
+ 176
+ ],
+ "score": 0.85,
+ "content": "y _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 164,
+ 378,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is reconstructed, or predicted rating, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 166,
+ 392,
+ 176
+ ],
+ "score": 0.87,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is a mask function such that",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 176,
+ 506,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 177,
+ 140,
+ 187
+ ],
+ "score": 0.91,
+ "content": "m _ { i } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 176,
+ 150,
+ 187
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 176,
+ 180,
+ 187
+ ],
+ "score": 0.91,
+ "content": "r _ { i } \\neq 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 176,
+ 199,
+ 187
+ ],
+ "score": 1.0,
+ "content": "else",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 177,
+ 232,
+ 187
+ ],
+ "score": 0.89,
+ "content": "m _ { i } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 176,
+ 506,
+ 187
+ ],
+ "score": 1.0,
+ "content": ". Note that there is a straightforward relation between RMSE score",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 277,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 182,
+ 199
+ ],
+ "score": 1.0,
+ "content": "and MMSE score:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 187,
+ 274,
+ 199
+ ],
+ "score": 0.9,
+ "content": "R M S E = \\sqrt { M M S E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 188,
+ 277,
+ 199
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 213,
+ 217,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 218,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 218,
+ 226
+ ],
+ "score": 1.0,
+ "content": "2.2 DENSE RE-FEEDING",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 233,
+ 505,
+ 311
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 265,
+ 246
+ ],
+ "score": 1.0,
+ "content": "During training and inference, an input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 265,
+ 234,
+ 298,
+ 244
+ ],
+ "score": 0.91,
+ "content": "x \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "is very sparse because no user can realistically rate",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "but a tiny fractions of all items. This poses problem for model training. Bayesian approches can",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 504,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 483,
+ 269
+ ],
+ "score": 1.0,
+ "content": "be used to overcome this issue (Wang et al., 2015). On the other hand, autoencoder’s output",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 255,
+ 504,
+ 267
+ ],
+ "score": 0.91,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 505,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 349,
+ 279
+ ],
+ "score": 1.0,
+ "content": "is dense. Lets consider an idealized scenario with a perfect",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 267,
+ 356,
+ 278
+ ],
+ "score": 0.83,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 266,
+ 385,
+ 279
+ ],
+ "score": 1.0,
+ "content": ". Then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 266,
+ 487,
+ 278
+ ],
+ "score": 0.91,
+ "content": "f ( x ) _ { i } = x _ { i } , \\forall i : x _ { i } \\neq 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 266,
+ 505,
+ 279
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 107,
+ 277,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 277,
+ 130,
+ 289
+ ],
+ "score": 0.91,
+ "content": "f ( { \\boldsymbol { x } } ) _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 277,
+ 343,
+ 290
+ ],
+ "score": 1.0,
+ "content": "accurately predicts all user’s future ratings for items",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 278,
+ 384,
+ 288
+ ],
+ "score": 0.91,
+ "content": "i : x _ { i } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 277,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": ". This means that if user rates",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 146,
+ 302
+ ],
+ "score": 1.0,
+ "content": "new item",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 289,
+ 153,
+ 298
+ ],
+ "score": 0.75,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 154,
+ 288,
+ 279,
+ 302
+ ],
+ "score": 1.0,
+ "content": "(thereby creating a new vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 289,
+ 289,
+ 299
+ ],
+ "score": 0.82,
+ "content": "x ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 288,
+ 314,
+ 302
+ ],
+ "score": 1.0,
+ "content": ") then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 288,
+ 365,
+ 300
+ ],
+ "score": 0.92,
+ "content": "f ( x ) _ { k } = x _ { k } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 288,
+ 384,
+ 302
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 288,
+ 441,
+ 300
+ ],
+ "score": 0.94,
+ "content": "f ( x ) = f ( x ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 441,
+ 288,
+ 506,
+ 302
+ ],
+ "score": 1.0,
+ "content": ". Hence, in this",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 481,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 183,
+ 313
+ ],
+ "score": 1.0,
+ "content": "idealized scenario,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 299,
+ 222,
+ 311
+ ],
+ "score": 0.93,
+ "content": "y = f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 298,
+ 438,
+ 313
+ ],
+ "score": 1.0,
+ "content": "should be a fixed point of a well trained autoencoder:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 299,
+ 477,
+ 311
+ ],
+ "score": 0.92,
+ "content": "f ( y ) = y",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 298,
+ 481,
+ 313
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 316,
+ 505,
+ 349
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 330
+ ],
+ "score": 1.0,
+ "content": "To explicitly enforce fixed-point constraint and to be able to perform dense training updates, we",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "augment every optimization iteration with an iterative dense re-feeding steps (3 and 4 below) as",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 337,
+ 142,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 142,
+ 350
+ ],
+ "score": 1.0,
+ "content": "follows:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 129,
+ 358,
+ 505,
+ 427
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 358,
+ 454,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 358,
+ 196,
+ 372
+ ],
+ "score": 1.0,
+ "content": "1. Given sparse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 362,
+ 203,
+ 369
+ ],
+ "score": 0.68,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 358,
+ 269,
+ 372
+ ],
+ "score": 1.0,
+ "content": ", compute dense",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 359,
+ 289,
+ 371
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 358,
+ 454,
+ 372
+ ],
+ "score": 1.0,
+ "content": "and loss using equation 1 (forward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 129,
+ 374,
+ 398,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 374,
+ 398,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2. Compute gradients and perform weight update (backward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 128,
+ 388,
+ 505,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 388,
+ 165,
+ 402
+ ],
+ "score": 1.0,
+ "content": "3. Treat",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 389,
+ 186,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 388,
+ 312,
+ 402
+ ],
+ "score": 1.0,
+ "content": "as a new example and compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 388,
+ 345,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( f ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 388,
+ 392,
+ 402
+ ],
+ "score": 1.0,
+ "content": ". Now both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 389,
+ 412,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 388,
+ 430,
+ 402
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 389,
+ 464,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( f ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 464,
+ 388,
+ 505,
+ 402
+ ],
+ "score": 1.0,
+ "content": "are dense",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 400,
+ 444,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 400,
+ 286,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and the loss from equation 1 has all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 402,
+ 296,
+ 410
+ ],
+ "score": 0.72,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 400,
+ 444,
+ 412
+ ],
+ "score": 1.0,
+ "content": "as non-zeros. (second forward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 129,
+ 415,
+ 428,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 415,
+ 428,
+ 428
+ ],
+ "score": 1.0,
+ "content": "4. Compute gradients and perform weight update (second backward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 435,
+ 406,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 406,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 406,
+ 448
+ ],
+ "score": 1.0,
+ "content": "Steps (3) and (4) can be also performed more than once for every iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 462,
+ 275,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 461,
+ 276,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 461,
+ 276,
+ 478
+ ],
+ "score": 1.0,
+ "content": "3 EXPERIMENTS AND RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 218,
+ 499
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 219,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 219,
+ 500
+ ],
+ "score": 1.0,
+ "content": "3.1 EXPERIMENT SETUP",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 505,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "For the rating prediction task, it is often most relevant to predict future ratings given the past ones",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "instead of predicting ratings missing at random. For evaluation purposes we followed Wu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "(2017) exactly by splitting the original Netflix Prize Bennett et al. (2007) training set into several",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "training and testing intervals based on time. Training interval contains ratings which came in earlier",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "than the ones from testing interval. Testing interval is then randomly split into Test and Validation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 329,
+ 576
+ ],
+ "score": 1.0,
+ "content": "subsets so that each rating from testing interval has a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 563,
+ 350,
+ 574
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "chance of appearing in either subset.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "Users and items that do not appear in the training set are removed from both test and validation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 309,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 309,
+ 597
+ ],
+ "score": 1.0,
+ "content": "subsets. Table 1 provides details on the data sets.2",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 602,
+ 505,
+ 646
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "For most of our experiments we uses a batch size of 128, trained using SGD with momentum of 0.9",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 625
+ ],
+ "score": 1.0,
+ "content": "and learning rate of 0.001. We used xavier initialization to initialize parameters. Note, that unlike",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Strub & Mary (2015) we did not use any layer-wise pre-training. We believe that we were able to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 450,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 450,
+ 648
+ ],
+ "score": 1.0,
+ "content": "do so successfully because of choosing the right activation function (see Section 3.2).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 660,
+ 286,
+ 671
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 288,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 288,
+ 672
+ ],
+ "score": 1.0,
+ "content": "3.2 EFFECTS OF THE ACTIVATION TYPES",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 39
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 680,
+ 504,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "To explore the effects of using different activation functions, we tested some of the most popular",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 691,
+ 505,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 691,
+ 386,
+ 704
+ ],
+ "score": 1.0,
+ "content": "choices in deep learning : sigmoid, “rectified linear units” (RELU),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 691,
+ 455,
+ 703
+ ],
+ "score": 0.9,
+ "content": "m a x ( r e l u ( x ) , 6 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 691,
+ 505,
+ 704
+ ],
+ "score": 1.0,
+ "content": "or RELU6,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 711,
+ 506,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "2Note, that while checking our data set statistics with first author of (Wu et al., 2017) it was determined that",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 414,
+ 731
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 393,
+ 731
+ ],
+ "score": 1.0,
+ "content": "their publication contained the following typo: “Netflix 6m” should be “Netflix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 393,
+ 722,
+ 409,
+ 731
+ ],
+ "score": 0.68,
+ "content": "3 \\mathrm { m } ^ { \\dag }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 410,
+ 721,
+ 414,
+ 731
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 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 2018",
+ "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": "title",
+ "bbox": [
+ 107,
+ 82,
+ 201,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 202,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 202,
+ 95
+ ],
+ "score": 1.0,
+ "content": "2.1 LOSS FUNCTION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 103,
+ 504,
+ 126
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 505,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 103,
+ 400,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Since it doesn’t make sense to predict zeros in user’s representation vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 400,
+ 106,
+ 407,
+ 113
+ ],
+ "score": 0.78,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 407,
+ 103,
+ 505,
+ 115
+ ],
+ "score": 1.0,
+ "content": ", we follow the approach",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 113,
+ 407,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 113,
+ 407,
+ 127
+ ],
+ "score": 1.0,
+ "content": "from Sedhain et al. (2015) and optimize Masked Mean Squared Error loss:",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1.5,
+ "bbox_fs": [
+ 105,
+ 103,
+ 505,
+ 127
+ ]
+ },
+ {
+ "type": "interline_equation",
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 161
+ ],
+ "score": 0.95,
+ "content": "M M S E = \\frac { m _ { i } * ( r _ { i } - y _ { i } ) ^ { 2 } } { \\sum _ { i = 0 } ^ { i = n } m _ { i } }",
+ "type": "interline_equation",
+ "image_path": "1897d4b1db1749e191b172c3dd7d98ae749d38dd07fd68a752082ec3404f525a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 247,
+ 130,
+ 364,
+ 145.5
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 247,
+ 145.5,
+ 364,
+ 161.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 164,
+ 504,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 164,
+ 133,
+ 177
+ ],
+ "score": 1.0,
+ "content": "where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 142,
+ 176
+ ],
+ "score": 0.86,
+ "content": "r _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 164,
+ 206,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is actual rating,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 206,
+ 167,
+ 215,
+ 176
+ ],
+ "score": 0.85,
+ "content": "y _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 164,
+ 378,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is reconstructed, or predicted rating, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 379,
+ 166,
+ 392,
+ 176
+ ],
+ "score": 0.87,
+ "content": "m _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 392,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "is a mask function such that",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 107,
+ 176,
+ 506,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 177,
+ 140,
+ 187
+ ],
+ "score": 0.91,
+ "content": "m _ { i } = 1",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 176,
+ 150,
+ 187
+ ],
+ "score": 1.0,
+ "content": "if",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 150,
+ 176,
+ 180,
+ 187
+ ],
+ "score": 0.91,
+ "content": "r _ { i } \\neq 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 180,
+ 176,
+ 199,
+ 187
+ ],
+ "score": 1.0,
+ "content": "else",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 199,
+ 177,
+ 232,
+ 187
+ ],
+ "score": 0.89,
+ "content": "m _ { i } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 176,
+ 506,
+ 187
+ ],
+ "score": 1.0,
+ "content": ". Note that there is a straightforward relation between RMSE score",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 187,
+ 277,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 182,
+ 199
+ ],
+ "score": 1.0,
+ "content": "and MMSE score:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 187,
+ 274,
+ 199
+ ],
+ "score": 0.9,
+ "content": "R M S E = \\sqrt { M M S E }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 274,
+ 188,
+ 277,
+ 199
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 164,
+ 506,
+ 199
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 213,
+ 217,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 218,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 213,
+ 218,
+ 226
+ ],
+ "score": 1.0,
+ "content": "2.2 DENSE RE-FEEDING",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 233,
+ 505,
+ 311
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 265,
+ 246
+ ],
+ "score": 1.0,
+ "content": "During training and inference, an input",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 265,
+ 234,
+ 298,
+ 244
+ ],
+ "score": 0.91,
+ "content": "x \\in R ^ { n }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 234,
+ 505,
+ 246
+ ],
+ "score": 1.0,
+ "content": "is very sparse because no user can realistically rate",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 245,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "but a tiny fractions of all items. This poses problem for model training. Bayesian approches can",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 255,
+ 504,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 483,
+ 269
+ ],
+ "score": 1.0,
+ "content": "be used to overcome this issue (Wang et al., 2015). On the other hand, autoencoder’s output",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 483,
+ 255,
+ 504,
+ 267
+ ],
+ "score": 0.91,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 505,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 349,
+ 279
+ ],
+ "score": 1.0,
+ "content": "is dense. Lets consider an idealized scenario with a perfect",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 349,
+ 267,
+ 356,
+ 278
+ ],
+ "score": 0.83,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 357,
+ 266,
+ 385,
+ 279
+ ],
+ "score": 1.0,
+ "content": ". Then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 266,
+ 487,
+ 278
+ ],
+ "score": 0.91,
+ "content": "f ( x ) _ { i } = x _ { i } , \\forall i : x _ { i } \\neq 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 266,
+ 505,
+ 279
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 107,
+ 277,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 277,
+ 130,
+ 289
+ ],
+ "score": 0.91,
+ "content": "f ( { \\boldsymbol { x } } ) _ { i }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 277,
+ 343,
+ 290
+ ],
+ "score": 1.0,
+ "content": "accurately predicts all user’s future ratings for items",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 343,
+ 278,
+ 384,
+ 288
+ ],
+ "score": 0.91,
+ "content": "i : x _ { i } = 0",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 385,
+ 277,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": ". This means that if user rates",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 146,
+ 302
+ ],
+ "score": 1.0,
+ "content": "new item",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 289,
+ 153,
+ 298
+ ],
+ "score": 0.75,
+ "content": "k",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 154,
+ 288,
+ 279,
+ 302
+ ],
+ "score": 1.0,
+ "content": "(thereby creating a new vector",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 280,
+ 289,
+ 289,
+ 299
+ ],
+ "score": 0.82,
+ "content": "x ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 288,
+ 314,
+ 302
+ ],
+ "score": 1.0,
+ "content": ") then",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 315,
+ 288,
+ 365,
+ 300
+ ],
+ "score": 0.92,
+ "content": "f ( x ) _ { k } = x _ { k } ^ { \\prime }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 288,
+ 384,
+ 302
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 384,
+ 288,
+ 441,
+ 300
+ ],
+ "score": 0.94,
+ "content": "f ( x ) = f ( x ^ { \\prime } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 441,
+ 288,
+ 506,
+ 302
+ ],
+ "score": 1.0,
+ "content": ". Hence, in this",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 481,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 183,
+ 313
+ ],
+ "score": 1.0,
+ "content": "idealized scenario,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 299,
+ 222,
+ 311
+ ],
+ "score": 0.93,
+ "content": "y = f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 222,
+ 298,
+ 438,
+ 313
+ ],
+ "score": 1.0,
+ "content": "should be a fixed point of a well trained autoencoder:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 438,
+ 299,
+ 477,
+ 311
+ ],
+ "score": 0.92,
+ "content": "f ( y ) = y",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 477,
+ 298,
+ 481,
+ 313
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 12,
+ "bbox_fs": [
+ 105,
+ 234,
+ 506,
+ 313
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 316,
+ 505,
+ 349
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 315,
+ 505,
+ 330
+ ],
+ "score": 1.0,
+ "content": "To explicitly enforce fixed-point constraint and to be able to perform dense training updates, we",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 328,
+ 505,
+ 339
+ ],
+ "score": 1.0,
+ "content": "augment every optimization iteration with an iterative dense re-feeding steps (3 and 4 below) as",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 337,
+ 142,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 337,
+ 142,
+ 350
+ ],
+ "score": 1.0,
+ "content": "follows:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 315,
+ 505,
+ 350
+ ]
+ },
+ {
+ "type": "index",
+ "bbox": [
+ 129,
+ 358,
+ 505,
+ 427
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 358,
+ 454,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 358,
+ 196,
+ 372
+ ],
+ "score": 1.0,
+ "content": "1. Given sparse",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 196,
+ 362,
+ 203,
+ 369
+ ],
+ "score": 0.68,
+ "content": "x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 358,
+ 269,
+ 372
+ ],
+ "score": 1.0,
+ "content": ", compute dense",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 359,
+ 289,
+ 371
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 358,
+ 454,
+ 372
+ ],
+ "score": 1.0,
+ "content": "and loss using equation 1 (forward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 129,
+ 374,
+ 398,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 374,
+ 398,
+ 386
+ ],
+ "score": 1.0,
+ "content": "2. Compute gradients and perform weight update (backward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 128,
+ 388,
+ 505,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 388,
+ 165,
+ 402
+ ],
+ "score": 1.0,
+ "content": "3. Treat",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 165,
+ 389,
+ 186,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 186,
+ 388,
+ 312,
+ 402
+ ],
+ "score": 1.0,
+ "content": "as a new example and compute",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 312,
+ 388,
+ 345,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( f ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 346,
+ 388,
+ 392,
+ 402
+ ],
+ "score": 1.0,
+ "content": ". Now both",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 392,
+ 389,
+ 412,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( x )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 412,
+ 388,
+ 430,
+ 402
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 430,
+ 389,
+ 464,
+ 401
+ ],
+ "score": 0.92,
+ "content": "f ( f ( x ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 464,
+ 388,
+ 505,
+ 402
+ ],
+ "score": 1.0,
+ "content": "are dense",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 141,
+ 400,
+ 444,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 400,
+ 286,
+ 412
+ ],
+ "score": 1.0,
+ "content": "and the loss from equation 1 has all",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 286,
+ 402,
+ 296,
+ 410
+ ],
+ "score": 0.72,
+ "content": "m",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 400,
+ 444,
+ 412
+ ],
+ "score": 1.0,
+ "content": "as non-zeros. (second forward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 22,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 129,
+ 415,
+ 428,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 415,
+ 428,
+ 428
+ ],
+ "score": 1.0,
+ "content": "4. Compute gradients and perform weight update (second backward pass)",
+ "type": "text"
+ }
+ ],
+ "index": 23,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 128,
+ 358,
+ 505,
+ 428
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 435,
+ 406,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 406,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 434,
+ 406,
+ 448
+ ],
+ "score": 1.0,
+ "content": "Steps (3) and (4) can be also performed more than once for every iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 106,
+ 434,
+ 406,
+ 448
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 462,
+ 275,
+ 475
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 461,
+ 276,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 461,
+ 276,
+ 478
+ ],
+ "score": 1.0,
+ "content": "3 EXPERIMENTS AND RESULTS",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 487,
+ 218,
+ 499
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 219,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 488,
+ 219,
+ 500
+ ],
+ "score": 1.0,
+ "content": "3.1 EXPERIMENT SETUP",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 508,
+ 505,
+ 596
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 509,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "For the rating prediction task, it is often most relevant to predict future ratings given the past ones",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 505,
+ 532
+ ],
+ "score": 1.0,
+ "content": "instead of predicting ratings missing at random. For evaluation purposes we followed Wu et al.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 531,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "(2017) exactly by splitting the original Netflix Prize Bennett et al. (2007) training set into several",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 541,
+ 505,
+ 554
+ ],
+ "score": 1.0,
+ "content": "training and testing intervals based on time. Training interval contains ratings which came in earlier",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "than the ones from testing interval. Testing interval is then randomly split into Test and Validation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 563,
+ 329,
+ 576
+ ],
+ "score": 1.0,
+ "content": "subsets so that each rating from testing interval has a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 330,
+ 563,
+ 350,
+ 574
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 350,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "chance of appearing in either subset.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 574,
+ 505,
+ 587
+ ],
+ "score": 1.0,
+ "content": "Users and items that do not appear in the training set are removed from both test and validation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 309,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 309,
+ 597
+ ],
+ "score": 1.0,
+ "content": "subsets. Table 1 provides details on the data sets.2",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 509,
+ 505,
+ 597
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 602,
+ 505,
+ 646
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 602,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "For most of our experiments we uses a batch size of 128, trained using SGD with momentum of 0.9",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 613,
+ 505,
+ 625
+ ],
+ "score": 1.0,
+ "content": "and learning rate of 0.001. We used xavier initialization to initialize parameters. Note, that unlike",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 623,
+ 506,
+ 637
+ ],
+ "score": 1.0,
+ "content": "Strub & Mary (2015) we did not use any layer-wise pre-training. We believe that we were able to",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 635,
+ 450,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 635,
+ 450,
+ 648
+ ],
+ "score": 1.0,
+ "content": "do so successfully because of choosing the right activation function (see Section 3.2).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 602,
+ 506,
+ 648
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 660,
+ 286,
+ 671
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 288,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 659,
+ 288,
+ 672
+ ],
+ "score": 1.0,
+ "content": "3.2 EFFECTS OF THE ACTIVATION TYPES",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 39
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 680,
+ 504,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 505,
+ 693
+ ],
+ "score": 1.0,
+ "content": "To explore the effects of using different activation functions, we tested some of the most popular",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 691,
+ 505,
+ 704
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 691,
+ 386,
+ 704
+ ],
+ "score": 1.0,
+ "content": "choices in deep learning : sigmoid, “rectified linear units” (RELU),",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 386,
+ 691,
+ 455,
+ 703
+ ],
+ "score": 0.9,
+ "content": "m a x ( r e l u ( x ) , 6 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 691,
+ 505,
+ 704
+ ],
+ "score": 1.0,
+ "content": "or RELU6,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "hyperbolic tangent (TANH), “exponential linear units” (ELU) (Clevert et al., 2015), leaky relu",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "(LRELU) (Xu et al., 2015) , “self-gated activation function” (SWISH) (Ramachandran et al., 2017),",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "and “scaled exponential linear units” (Klambauer et al., 2017) (SELU) on the 4 layer autoencoder",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "with 128 units in each hidden layer. Because ratings are on the scale from 1 to 5, we keep last layer",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "of the decoder linear for sigmoid and tanh-based models. In all other models activation function is",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 190,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 190,
+ 662
+ ],
+ "score": 1.0,
+ "content": "applied in all layers.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 106,
+ 680,
+ 505,
+ 704
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "09a17f90285f07af06cb356344fc9c7e4c66332e57baa5666e3e4a13acf00cc6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 157.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 130,
+ 157.33333333333331,
+ 482,
+ 232.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 130,
+ 232.66666666666663,
+ 482,
+ 307.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 359
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Figure 1: AutoEncoder consists of two neural networks, encoder and decoder, fused together on",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 213,
+ 349
+ ],
+ "score": 1.0,
+ "content": "the “representation” layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 338,
+ 220,
+ 346
+ ],
+ "score": 0.67,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 336,
+ 313,
+ 349
+ ],
+ "score": 1.0,
+ "content": ". Encoder has 2 layers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 338,
+ 324,
+ 347
+ ],
+ "score": 0.84,
+ "content": "e _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 336,
+ 343,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 338,
+ 354,
+ 347
+ ],
+ "score": 0.84,
+ "content": "e _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 336,
+ 460,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and decoder has 2 layers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 337,
+ 471,
+ 347
+ ],
+ "score": 0.88,
+ "content": "d _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 336,
+ 490,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 336,
+ 501,
+ 347
+ ],
+ "score": 0.86,
+ "content": "d _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 336,
+ 505,
+ 349
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 279,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 267,
+ 360
+ ],
+ "score": 1.0,
+ "content": "Dropout may be applied to coding layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 349,
+ 274,
+ 357
+ ],
+ "score": 0.72,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 347,
+ 279,
+ 360
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 421
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Table 1: Subsets of Netflix Prize training set used in our experiments. We made sure that these splits",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 409,
+ 303,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 303,
+ 421
+ ],
+ "score": 1.0,
+ "content": "match exactly the ones used in (Wu et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "score": 0.981,
+ "html": "
| Full | 3 months | 6 months | 1 year |
| | | 06/05-11/05 | |
| Training Users | 12/99-11/05 477,412 | 09/05-11/05 311,315 | 390,795 | 06/04-05/05 345,855 |
| Ratings | 98,074,901 | 13,675,402 | 29,179,009 | 41,451,832 |
| | | | |
| Testing Users | 12/05 | 12/05 | 12/05 | 06/05 |
",
+ "type": "table",
+ "image_path": "98febd8b2de97993caeb17fa47cd2eb95b84983871e8e9652b9662c8b7406a28.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 462.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 159,
+ 462.0,
+ 452,
+ 500.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 159,
+ 500.0,
+ 452,
+ 538.0
+ ],
+ "spans": [],
+ "index": 10
+ }
+ ]
+ }
+ ],
+ "index": 7.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "hyperbolic tangent (TANH), “exponential linear units” (ELU) (Clevert et al., 2015), leaky relu",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "(LRELU) (Xu et al., 2015) , “self-gated activation function” (SWISH) (Ramachandran et al., 2017),",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 616,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 616,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "and “scaled exponential linear units” (Klambauer et al., 2017) (SELU) on the 4 layer autoencoder",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "with 128 units in each hidden layer. Because ratings are on the scale from 1 to 5, we keep last layer",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "of the decoder linear for sigmoid and tanh-based models. In all other models activation function is",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 649,
+ 190,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 190,
+ 662
+ ],
+ "score": 1.0,
+ "content": "applied in all layers.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 665,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "We found that on this task ELU, SELU and LRELU perform much better than SIGMOID, RELU,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "RELU6, TANH and SWISH. Figure 2 clearly demonstrates this. There are two properties which",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "seems to separate activations which perform well from those which do not: a) non-zero negative",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "part and b) unbounded positive part. Hence, we conclude, that in this setting these properties are im-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "portant for successful training. Thus, we use SELU activation units and tune SELU-based networks",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 176,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 176,
+ 734
+ ],
+ "score": 1.0,
+ "content": "for performance.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 19.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 308
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "09a17f90285f07af06cb356344fc9c7e4c66332e57baa5666e3e4a13acf00cc6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 130,
+ 82,
+ 482,
+ 157.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 130,
+ 157.33333333333331,
+ 482,
+ 232.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 130,
+ 232.66666666666663,
+ 482,
+ 307.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 325,
+ 505,
+ 359
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Figure 1: AutoEncoder consists of two neural networks, encoder and decoder, fused together on",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 213,
+ 349
+ ],
+ "score": 1.0,
+ "content": "the “representation” layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 213,
+ 338,
+ 220,
+ 346
+ ],
+ "score": 0.67,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 220,
+ 336,
+ 313,
+ 349
+ ],
+ "score": 1.0,
+ "content": ". Encoder has 2 layers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 314,
+ 338,
+ 324,
+ 347
+ ],
+ "score": 0.84,
+ "content": "e _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 325,
+ 336,
+ 343,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 344,
+ 338,
+ 354,
+ 347
+ ],
+ "score": 0.84,
+ "content": "e _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 355,
+ 336,
+ 460,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and decoder has 2 layers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 460,
+ 337,
+ 471,
+ 347
+ ],
+ "score": 0.88,
+ "content": "d _ { 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 471,
+ 336,
+ 490,
+ 349
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 336,
+ 501,
+ 347
+ ],
+ "score": 0.86,
+ "content": "d _ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 501,
+ 336,
+ 505,
+ 349
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 347,
+ 279,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 347,
+ 267,
+ 360
+ ],
+ "score": 1.0,
+ "content": "Dropout may be applied to coding layer",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 268,
+ 349,
+ 274,
+ 357
+ ],
+ "score": 0.72,
+ "content": "z",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 347,
+ 279,
+ 360
+ ],
+ "score": 1.0,
+ "content": ".",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 421
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Table 1: Subsets of Netflix Prize training set used in our experiments. We made sure that these splits",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 409,
+ 303,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 409,
+ 303,
+ 421
+ ],
+ "score": 1.0,
+ "content": "match exactly the ones used in (Wu et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 538
+ ],
+ "score": 0.981,
+ "html": " | Full | 3 months | 6 months | 1 year |
| | | 06/05-11/05 | |
| Training Users | 12/99-11/05 477,412 | 09/05-11/05 311,315 | 390,795 | 06/04-05/05 345,855 |
| Ratings | 98,074,901 | 13,675,402 | 29,179,009 | 41,451,832 |
| | | | |
| Testing Users | 12/05 | 12/05 | 12/05 | 06/05 |
",
+ "type": "table",
+ "image_path": "98febd8b2de97993caeb17fa47cd2eb95b84983871e8e9652b9662c8b7406a28.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 159,
+ 424,
+ 452,
+ 462.0
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 159,
+ 462.0,
+ 452,
+ 500.0
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 159,
+ 500.0,
+ 452,
+ 538.0
+ ],
+ "spans": [],
+ "index": 10
+ }
+ ]
+ }
+ ],
+ "index": 7.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 593,
+ 505,
+ 660
+ ],
+ "lines": [],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 594,
+ 506,
+ 662
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 665,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 666,
+ 505,
+ 677
+ ],
+ "score": 1.0,
+ "content": "We found that on this task ELU, SELU and LRELU perform much better than SIGMOID, RELU,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "RELU6, TANH and SWISH. Figure 2 clearly demonstrates this. There are two properties which",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "seems to separate activations which perform well from those which do not: a) non-zero negative",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "part and b) unbounded positive part. Hence, we conclude, that in this setting these properties are im-",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "portant for successful training. Thus, we use SELU activation units and tune SELU-based networks",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 176,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 176,
+ 734
+ ],
+ "score": 1.0,
+ "content": "for performance.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 666,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "aec9a2eced5f723ac04e892d3f29d7e2ede096781d457488e2eaddd2ec164ecb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 148.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 148.0,
+ 500,
+ 212.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 212.0,
+ 500,
+ 276.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 338
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Training RMSE per mini-batch. All lines correspond to 4-layers autoencoder (2 layer",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "encoder and 2 layer decoder) with hidden unit dimensions of 128. Different line colors correspond",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "to different activation functions. TANH and SIGMOID lines are very similar as well as lines for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 424,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 424,
+ 339
+ ],
+ "score": 1.0,
+ "content": "ELU and SELU. The best performing activation functions are ELU and SELU.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 358,
+ 238,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 239,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 239,
+ 370
+ ],
+ "score": 1.0,
+ "content": "3.3 OVER-FITTING THE DATA",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 379,
+ 504,
+ 413
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "The largest data set we use for training, “Netflix Full” from Table 1, contains 98M ratings given by",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 337,
+ 403
+ ],
+ "score": 1.0,
+ "content": "477K users. Number of movies (e.g. items) in this set is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 391,
+ 369,
+ 401
+ ],
+ "score": 0.76,
+ "content": "n = 1 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 389,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": ", 768. Therefore, the first layer of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 400,
+ 415,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 400,
+ 179,
+ 415
+ ],
+ "score": 1.0,
+ "content": "encoder will have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 402,
+ 219,
+ 412
+ ],
+ "score": 0.91,
+ "content": "d * n + d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 400,
+ 282,
+ 415
+ ],
+ "score": 1.0,
+ "content": "weights, where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 402,
+ 289,
+ 411
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 400,
+ 415,
+ 415
+ ],
+ "score": 1.0,
+ "content": "is number of units in the layer.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "For modern deep learning algorithms and hardware this is relatively small task. If we start with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 430,
+ 505,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 430,
+ 451,
+ 442
+ ],
+ "score": 1.0,
+ "content": "single layer encoders and decoders we can quickly overfit to the training data even for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 430,
+ 459,
+ 439
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 459,
+ 430,
+ 505,
+ 442
+ ],
+ "score": 1.0,
+ "content": "as small as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "512. Figure 3 clearly demonstrates this. Switching from unconstrained autoencoder to constrained",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 362,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 362,
+ 463
+ ],
+ "score": 1.0,
+ "content": "reduces over-fitting, but does not completely solve the problem.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "63b7bb1ea66cdf6bb9ca0bb346e6100321c6a5f1f9ea9693554e275cb0692ba0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 516.3333333333334
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 516.3333333333334,
+ 505,
+ 557.6666666666667
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 557.6666666666667,
+ 505,
+ 599.0000000000001
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 613,
+ 504,
+ 636
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 504,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 504,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Single layer autoencoder with 128, 256, 512 and 1024 hidden units in the coding layer. A:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 342,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 342,
+ 638
+ ],
+ "score": 1.0,
+ "content": "training RMSE per epoch; B: evaluation RMSE per epoch.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 656,
+ 196,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 198,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 198,
+ 668
+ ],
+ "score": 1.0,
+ "content": "3.4 GOING DEEPER",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 676,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "While making layers wider helps bring training loss down, adding more layers is often correlated",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 686,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "with a network’s ability to generalize. In this set of experiments we show that this is indeed the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 318,
+ 711
+ ],
+ "score": 1.0,
+ "content": "case here. We choose small enough dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 699,
+ 356,
+ 709
+ ],
+ "score": 0.85,
+ "content": "d = 1 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": ") for all hidden layers to easily avoid",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "over-fitting and start adding more layers. Table 2 shows that there is a positive correlation between",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 305,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 305,
+ 734
+ ],
+ "score": 1.0,
+ "content": "the number of layers and the evaluation accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "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"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 25,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 25,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 276
+ ],
+ "score": 0.96,
+ "type": "image",
+ "image_path": "aec9a2eced5f723ac04e892d3f29d7e2ede096781d457488e2eaddd2ec164ecb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 84,
+ 500,
+ 148.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 148.0,
+ 500,
+ 212.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 212.0,
+ 500,
+ 276.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 293,
+ 505,
+ 338
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Training RMSE per mini-batch. All lines correspond to 4-layers autoencoder (2 layer",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 304,
+ 506,
+ 317
+ ],
+ "score": 1.0,
+ "content": "encoder and 2 layer decoder) with hidden unit dimensions of 128. Different line colors correspond",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 316,
+ 506,
+ 328
+ ],
+ "score": 1.0,
+ "content": "to different activation functions. TANH and SIGMOID lines are very similar as well as lines for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 326,
+ 424,
+ 339
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 326,
+ 424,
+ 339
+ ],
+ "score": 1.0,
+ "content": "ELU and SELU. The best performing activation functions are ELU and SELU.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 358,
+ 238,
+ 370
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 239,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 358,
+ 239,
+ 370
+ ],
+ "score": 1.0,
+ "content": "3.3 OVER-FITTING THE DATA",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 379,
+ 504,
+ 413
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 505,
+ 392
+ ],
+ "score": 1.0,
+ "content": "The largest data set we use for training, “Netflix Full” from Table 1, contains 98M ratings given by",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 389,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 337,
+ 403
+ ],
+ "score": 1.0,
+ "content": "477K users. Number of movies (e.g. items) in this set is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 337,
+ 391,
+ 369,
+ 401
+ ],
+ "score": 0.76,
+ "content": "n = 1 7",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 369,
+ 389,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": ", 768. Therefore, the first layer of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 400,
+ 415,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 400,
+ 179,
+ 415
+ ],
+ "score": 1.0,
+ "content": "encoder will have",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 402,
+ 219,
+ 412
+ ],
+ "score": 0.91,
+ "content": "d * n + d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 219,
+ 400,
+ 282,
+ 415
+ ],
+ "score": 1.0,
+ "content": "weights, where",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 283,
+ 402,
+ 289,
+ 411
+ ],
+ "score": 0.77,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 400,
+ 415,
+ 415
+ ],
+ "score": 1.0,
+ "content": "is number of units in the layer.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 378,
+ 506,
+ 415
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 418,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 418,
+ 505,
+ 430
+ ],
+ "score": 1.0,
+ "content": "For modern deep learning algorithms and hardware this is relatively small task. If we start with",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 430,
+ 505,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 430,
+ 451,
+ 442
+ ],
+ "score": 1.0,
+ "content": "single layer encoders and decoders we can quickly overfit to the training data even for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 452,
+ 430,
+ 459,
+ 439
+ ],
+ "score": 0.79,
+ "content": "d",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 459,
+ 430,
+ 505,
+ 442
+ ],
+ "score": 1.0,
+ "content": "as small as",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "512. Figure 3 clearly demonstrates this. Switching from unconstrained autoencoder to constrained",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 452,
+ 362,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 362,
+ 463
+ ],
+ "score": 1.0,
+ "content": "reduces over-fitting, but does not completely solve the problem.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 106,
+ 418,
+ 505,
+ 463
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 599
+ ],
+ "score": 0.961,
+ "type": "image",
+ "image_path": "63b7bb1ea66cdf6bb9ca0bb346e6100321c6a5f1f9ea9693554e275cb0692ba0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 475,
+ 505,
+ 516.3333333333334
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 516.3333333333334,
+ 505,
+ 557.6666666666667
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 557.6666666666667,
+ 505,
+ 599.0000000000001
+ ],
+ "spans": [],
+ "index": 17
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 613,
+ 504,
+ 636
+ ],
+ "group_id": 1,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 504,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 613,
+ 504,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Single layer autoencoder with 128, 256, 512 and 1024 hidden units in the coding layer. A:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 624,
+ 342,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 624,
+ 342,
+ 638
+ ],
+ "score": 1.0,
+ "content": "training RMSE per epoch; B: evaluation RMSE per epoch.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "index": 17.25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 656,
+ 196,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 198,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 655,
+ 198,
+ 668
+ ],
+ "score": 1.0,
+ "content": "3.4 GOING DEEPER",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 676,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "While making layers wider helps bring training loss down, adding more layers is often correlated",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 686,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 686,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "with a network’s ability to generalize. In this set of experiments we show that this is indeed the",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 318,
+ 711
+ ],
+ "score": 1.0,
+ "content": "case here. We choose small enough dimensionality",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 699,
+ 356,
+ 709
+ ],
+ "score": 0.85,
+ "content": "d = 1 2 8",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 356,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": ") for all hidden layers to easily avoid",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 708,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "over-fitting and start adding more layers. Table 2 shows that there is a positive correlation between",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 720,
+ 305,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 720,
+ 305,
+ 734
+ ],
+ "score": 1.0,
+ "content": "the number of layers and the evaluation accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 677,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 111
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 101
+ ],
+ "score": 1.0,
+ "content": "Table 2: Depth helps generalization. Evaluation RMSE of the models with different number of",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 100,
+ 322,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 322,
+ 111
+ ],
+ "score": 1.0,
+ "content": "layers. In all cases the hidden layer dimension is 128.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "score": 0.975,
+ "html": "| Number of layers | Evaluation RMSE | params |
| | |
| 2 | 1.146 | 4,566,504 |
| 4 | 0.9615 | 4,599,528 |
| 6 | 0.9378 | 4,632,552 |
| 8 10 | 0.9364 | 4,665,576 |
| 12 | 0.9340 0.9328 | 4,698,600 |
| | 4,731,624 |
",
+ "type": "table",
+ "image_path": "88fcfcca1263e20a032f9acfaef74ea8491daebbea86756b7bd10814e85b9e95.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 129.57142857142858
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 190,
+ 129.57142857142858,
+ 421,
+ 144.14285714285717
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 190,
+ 144.14285714285717,
+ 421,
+ 158.71428571428575
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 190,
+ 158.71428571428575,
+ 421,
+ 173.28571428571433
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 190,
+ 173.28571428571433,
+ 421,
+ 187.85714285714292
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 190,
+ 187.85714285714292,
+ 421,
+ 202.4285714285715
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 190,
+ 202.4285714285715,
+ 421,
+ 217.00000000000009
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 309
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Going from one layer in encoder and decoder to three layers in both provides good improvement in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "evaluation RMSE (from 1.146 to 0.9378). After that, blindly adding more layers does help, however",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 376,
+ 289
+ ],
+ "score": 1.0,
+ "content": "it provides diminishing returns. Note that the model with single",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 276,
+ 416,
+ 286
+ ],
+ "score": 0.89,
+ "content": "d \\ : = \\ : 2 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 275,
+ 505,
+ 289
+ ],
+ "score": 1.0,
+ "content": "layer in encoder and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "decoder has 9,115,240 parameters which is almost two times more than any of these deep models",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 330,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 330,
+ 310
+ ],
+ "score": 1.0,
+ "content": "while having much worse evauation RMSE (above 1.0).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 323,
+ 174,
+ 335
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 176,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 176,
+ 336
+ ],
+ "score": 1.0,
+ "content": "3.5 DROPOUT",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 344,
+ 505,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Section 3.4 shows us that adding too many small layers eventually hits diminishing returns. Thus, we",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "start experimenting with model architecture and hyper-parameters more broadly. Our most promis-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 505,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 279,
+ 379
+ ],
+ "score": 1.0,
+ "content": "ing model has the following architecture:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 366,
+ 399,
+ 378
+ ],
+ "score": 0.59,
+ "content": "n , 5 1 2 , 5 1 2 , 1 0 2 4 , 5 1 2 , 5 1 2 , n .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 366,
+ 505,
+ 379
+ ],
+ "score": 1.0,
+ "content": ", which means 3 layers in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 377,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "encoder (512,512,1024), coding layer of 1024 and 3 layers in decoder of size 512,512,n. This",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "model, however, quickly over-fits if trained with no regularization. To regularize it, we tried several",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "dropout values and, interestingly, very high values of drop probability (e.g. 0.8) turned out to be",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 408,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 408,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "the best. See Figure 4 for evaluation RMSE. We apply dropout on the encoder output only, e.g.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 421,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 421,
+ 263,
+ 433
+ ],
+ "score": 0.86,
+ "content": "f ( x ) = d e c o d e ( d r o p o u t ( e n c o d e ( x ) ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 421,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": ". We tried applying dropout after every layer of the model",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 395,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 395,
+ 445
+ ],
+ "score": 1.0,
+ "content": "but that stifled training convergence and did not improve generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "1d42da41c3ad8938ff124090a7abb4a7d64268bb5591f52f0509568c13480560.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 496.0
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 182,
+ 496.0,
+ 428,
+ 538.0
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 182,
+ 538.0,
+ 428,
+ 580.0
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 640
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Effects of dropout. Y-axis: evaluation RMSE, X-axis: epoch number. Model with no",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "dropout (Drop Prob 0.0) clearly over-fits. Model with drop probability of 0.5 over-fits as well (but",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "much slowly). Models with drop probabilities of 0.65 and 0.8 result in RMSEs of 0.9192 and 0.9183",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 176,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 176,
+ 643
+ ],
+ "score": 1.0,
+ "content": "correspondingly.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ }
+ ],
+ "index": 26.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 667,
+ 216,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 217,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 217,
+ 678
+ ],
+ "score": 1.0,
+ "content": "3.6 DENSE RE-FEEDING",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "score": 1.0,
+ "content": "Iterative dense re-feeding (see Section 2.2) provides us with additional improvement in evaluation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 237,
+ 711
+ ],
+ "score": 1.0,
+ "content": "accuracy for our 6-layer-model:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 700,
+ 245,
+ 709
+ ],
+ "score": 0.44,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 699,
+ 311,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", 512, 512, 1024,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 699,
+ 343,
+ 711
+ ],
+ "score": 0.48,
+ "content": "d p ( 0 . 8 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 699,
+ 384,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", 512, 512,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 700,
+ 392,
+ 709
+ ],
+ "score": 0.64,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(referred to as Baseline be-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 455,
+ 723
+ ],
+ "score": 1.0,
+ "content": "low). Here each parameter denotes the number of inputs, hidden units, or outputs and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 709,
+ 487,
+ 721
+ ],
+ "score": 0.86,
+ "content": "d p ( 0 . 8 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "dropout layer with a drop probability of 0.8. Just applying output re-feeding did not have significant",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 33.5
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "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": "table",
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 111
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 101
+ ],
+ "score": 1.0,
+ "content": "Table 2: Depth helps generalization. Evaluation RMSE of the models with different number of",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 100,
+ 322,
+ 111
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 100,
+ 322,
+ 111
+ ],
+ "score": 1.0,
+ "content": "layers. In all cases the hidden layer dimension is 128.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 217
+ ],
+ "score": 0.975,
+ "html": "| Number of layers | Evaluation RMSE | params |
| | |
| 2 | 1.146 | 4,566,504 |
| 4 | 0.9615 | 4,599,528 |
| 6 | 0.9378 | 4,632,552 |
| 8 10 | 0.9364 | 4,665,576 |
| 12 | 0.9340 0.9328 | 4,698,600 |
| | 4,731,624 |
",
+ "type": "table",
+ "image_path": "88fcfcca1263e20a032f9acfaef74ea8491daebbea86756b7bd10814e85b9e95.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 190,
+ 115,
+ 421,
+ 129.57142857142858
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 190,
+ 129.57142857142858,
+ 421,
+ 144.14285714285717
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 190,
+ 144.14285714285717,
+ 421,
+ 158.71428571428575
+ ],
+ "spans": [],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 190,
+ 158.71428571428575,
+ 421,
+ 173.28571428571433
+ ],
+ "spans": [],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 190,
+ 173.28571428571433,
+ 421,
+ 187.85714285714292
+ ],
+ "spans": [],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 190,
+ 187.85714285714292,
+ 421,
+ 202.4285714285715
+ ],
+ "spans": [],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 190,
+ 202.4285714285715,
+ 421,
+ 217.00000000000009
+ ],
+ "spans": [],
+ "index": 8
+ }
+ ]
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 254,
+ 505,
+ 309
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Going from one layer in encoder and decoder to three layers in both provides good improvement in",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "evaluation RMSE (from 1.146 to 0.9378). After that, blindly adding more layers does help, however",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 275,
+ 505,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 275,
+ 376,
+ 289
+ ],
+ "score": 1.0,
+ "content": "it provides diminishing returns. Note that the model with single",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 276,
+ 416,
+ 286
+ ],
+ "score": 0.89,
+ "content": "d \\ : = \\ : 2 5 6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 416,
+ 275,
+ 505,
+ 289
+ ],
+ "score": 1.0,
+ "content": "layer in encoder and",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 286,
+ 505,
+ 300
+ ],
+ "score": 1.0,
+ "content": "decoder has 9,115,240 parameters which is almost two times more than any of these deep models",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 298,
+ 330,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 298,
+ 330,
+ 310
+ ],
+ "score": 1.0,
+ "content": "while having much worse evauation RMSE (above 1.0).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 105,
+ 254,
+ 505,
+ 310
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 323,
+ 174,
+ 335
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 176,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 176,
+ 336
+ ],
+ "score": 1.0,
+ "content": "3.5 DROPOUT",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 344,
+ 505,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Section 3.4 shows us that adding too many small layers eventually hits diminishing returns. Thus, we",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "start experimenting with model architecture and hyper-parameters more broadly. Our most promis-",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 366,
+ 505,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 366,
+ 279,
+ 379
+ ],
+ "score": 1.0,
+ "content": "ing model has the following architecture:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 279,
+ 366,
+ 399,
+ 378
+ ],
+ "score": 0.59,
+ "content": "n , 5 1 2 , 5 1 2 , 1 0 2 4 , 5 1 2 , 5 1 2 , n .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 399,
+ 366,
+ 505,
+ 379
+ ],
+ "score": 1.0,
+ "content": ", which means 3 layers in",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 377,
+ 506,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 377,
+ 506,
+ 389
+ ],
+ "score": 1.0,
+ "content": "encoder (512,512,1024), coding layer of 1024 and 3 layers in decoder of size 512,512,n. This",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 388,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 388,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "model, however, quickly over-fits if trained with no regularization. To regularize it, we tried several",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 505,
+ 412
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 505,
+ 412
+ ],
+ "score": 1.0,
+ "content": "dropout values and, interestingly, very high values of drop probability (e.g. 0.8) turned out to be",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 408,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 408,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "the best. See Figure 4 for evaluation RMSE. We apply dropout on the encoder output only, e.g.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 421,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 421,
+ 263,
+ 433
+ ],
+ "score": 0.86,
+ "content": "f ( x ) = d e c o d e ( d r o p o u t ( e n c o d e ( x ) ) )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 421,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": ". We tried applying dropout after every layer of the model",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 432,
+ 395,
+ 445
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 432,
+ 395,
+ 445
+ ],
+ "score": 1.0,
+ "content": "but that stifled training convergence and did not improve generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 104,
+ 343,
+ 506,
+ 445
+ ]
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 580
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "1d42da41c3ad8938ff124090a7abb4a7d64268bb5591f52f0509568c13480560.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 182,
+ 454,
+ 428,
+ 496.0
+ ],
+ "spans": [],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 182,
+ 496.0,
+ 428,
+ 538.0
+ ],
+ "spans": [],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 182,
+ 538.0,
+ 428,
+ 580.0
+ ],
+ "spans": [],
+ "index": 26
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 595,
+ 506,
+ 640
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 595,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Effects of dropout. Y-axis: evaluation RMSE, X-axis: epoch number. Model with no",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "dropout (Drop Prob 0.0) clearly over-fits. Model with drop probability of 0.5 over-fits as well (but",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "much slowly). Models with drop probabilities of 0.65 and 0.8 result in RMSEs of 0.9192 and 0.9183",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 629,
+ 176,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 629,
+ 176,
+ 643
+ ],
+ "score": 1.0,
+ "content": "correspondingly.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28.5
+ }
+ ],
+ "index": 26.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 667,
+ 216,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 217,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 667,
+ 217,
+ 678
+ ],
+ "score": 1.0,
+ "content": "3.6 DENSE RE-FEEDING",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 505,
+ 699
+ ],
+ "score": 1.0,
+ "content": "Iterative dense re-feeding (see Section 2.2) provides us with additional improvement in evaluation",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 237,
+ 711
+ ],
+ "score": 1.0,
+ "content": "accuracy for our 6-layer-model:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 237,
+ 700,
+ 245,
+ 709
+ ],
+ "score": 0.44,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 245,
+ 699,
+ 311,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", 512, 512, 1024,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 311,
+ 699,
+ 343,
+ 711
+ ],
+ "score": 0.48,
+ "content": "d p ( 0 . 8 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 699,
+ 384,
+ 711
+ ],
+ "score": 1.0,
+ "content": ", 512, 512,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 385,
+ 700,
+ 392,
+ 709
+ ],
+ "score": 0.64,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 393,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "(referred to as Baseline be-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 455,
+ 723
+ ],
+ "score": 1.0,
+ "content": "low). Here each parameter denotes the number of inputs, hidden units, or outputs and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 455,
+ 709,
+ 487,
+ 721
+ ],
+ "score": 0.86,
+ "content": "d p ( 0 . 8 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 487,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "is a",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "dropout layer with a drop probability of 0.8. Just applying output re-feeding did not have significant",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "impact on the model performance. However, in conjunction with the higher learning rate, it did",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "significantly increase the model performance. Note, that with this higher learning rate (0.005) but",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 424,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 424,
+ 253
+ ],
+ "score": 1.0,
+ "content": "without dense re-feeding, the model started to diverge. See Figure 5 for details.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 688,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 105,
+ 89,
+ 504,
+ 111
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 3: Test RMSE of different models. I-AR, U-AR and RRN numbers are taken from (Wu et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 98,
+ 134,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 98,
+ 134,
+ 113
+ ],
+ "score": 1.0,
+ "content": "2017)",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "score": 0.974,
+ "html": "| DataSet | I-AR | U-AR | RRN | DeepRec |
| Netflix 3 months Netfix Full | | | | |
| 0.9778 | 0.9836 | 0.9427 | 0.9373 |
| 0.9364 | 0.9647 | 0.9224 | 0.9099 |
",
+ "type": "table",
+ "image_path": "89ef14c3de00dd1d762baf07ab626bac7c2f3f10836603e833ad767ad6532fab.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 133.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 180,
+ 133.66666666666666,
+ 430,
+ 153.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 180,
+ 153.33333333333331,
+ 430,
+ 172.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 218,
+ 505,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "impact on the model performance. However, in conjunction with the higher learning rate, it did",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 230,
+ 505,
+ 242
+ ],
+ "score": 1.0,
+ "content": "significantly increase the model performance. Note, that with this higher learning rate (0.005) but",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 241,
+ 424,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 241,
+ 424,
+ 253
+ ],
+ "score": 1.0,
+ "content": "without dense re-feeding, the model started to diverge. See Figure 5 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "b589cd81bc68f7c0374d34eb5d18286cf197a9628be19ee2e0751e9faaaf5554.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 286.72727272727275
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 183,
+ 286.72727272727275,
+ 425,
+ 300.4545454545455
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 183,
+ 300.4545454545455,
+ 425,
+ 314.18181818181824
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 183,
+ 314.18181818181824,
+ 425,
+ 327.909090909091
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 183,
+ 327.909090909091,
+ 425,
+ 341.63636363636374
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 183,
+ 341.63636363636374,
+ 425,
+ 355.3636363636365
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 183,
+ 355.3636363636365,
+ 425,
+ 369.09090909090924
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 183,
+ 369.09090909090924,
+ 425,
+ 382.818181818182
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 183,
+ 382.818181818182,
+ 425,
+ 396.54545454545473
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 183,
+ 396.54545454545473,
+ 425,
+ 410.2727272727275
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 183,
+ 410.2727272727275,
+ 425,
+ 424.0000000000002
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 496
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 371,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Effects of dense re-feeding. Y-axis: evaluation RMSE,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 441,
+ 380,
+ 451
+ ],
+ "score": 0.28,
+ "content": "\\mathbf { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 380,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "-axis: epoch number. Baseline",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "model was trained with learning rate of 0.001. Applying re-feeding step with the same learning",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 476
+ ],
+ "score": 1.0,
+ "content": "rate almost did not help (Baseline RF). Learning rate of 0.005 (Baseline LR 0.005) is too big for",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "baseline model without re-feeding. However, increasing both learning rate and applying re-feeding",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 279,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 279,
+ 498
+ ],
+ "score": 1.0,
+ "content": "step clearly helps (Baseline LR 0.005 RF).",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 523,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Applying dense re-feeding and increasing the learning rate, allowed us to further improve the evalua-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "tion RMSE from 0.9167 to 0.9100. Picking a checkpoint with best evaluation RMSE and computing",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 459,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 459,
+ 558
+ ],
+ "score": 1.0,
+ "content": "test RMSE gives as 0.9099, which we believe is significantly better than other methods.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 580,
+ 290,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 291,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 291,
+ 592
+ ],
+ "score": 1.0,
+ "content": "3.7 COMPARISON WITH OTHER METHODS",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We compare our best model with Recurrent Recommender Network from Wu et al. (2017) which",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "has been shown to outperform PMF (Mnih & Salakhutdinov, 2008), T-SVD (Koren, 2010) and I/U-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "AR (Sedhain et al., 2015) on the data we use (see Table 1 for data description). Note, that unlike",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "T-SVD and RRN, our method does not explicitly take into account temporal dynamics of ratings.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Yet, Table 3 shows that it is still capable of outperforming these methods on future rating prediction",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "task. We train each model using only the training set and compute evaluation RMSE for 100 epochs.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 424,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 424,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Then the checkpoint with the highest evaluation RMSE is tested on the test set.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "“Netflix 3 months” has 7 times less training data compared to “Netflix full”, it is therefore, not",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "surprising that the model’s performance is significantly worse if trained on this data alone (0.9373",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "vs 0.9099). In fact, the model that performs best on “Netflix full” over-fits on this set, and we had",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 390,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 390,
+ 732
+ ],
+ "score": 1.0,
+ "content": "to reduce the model’s complexity accordingly (see Table 4 for details).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "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": "table",
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 105,
+ 89,
+ 504,
+ 111
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 88,
+ 505,
+ 102
+ ],
+ "score": 1.0,
+ "content": "Table 3: Test RMSE of different models. I-AR, U-AR and RRN numbers are taken from (Wu et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 98,
+ 134,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 98,
+ 134,
+ 113
+ ],
+ "score": 1.0,
+ "content": "2017)",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 173
+ ],
+ "score": 0.974,
+ "html": "| DataSet | I-AR | U-AR | RRN | DeepRec |
| Netflix 3 months Netfix Full | | | | |
| 0.9778 | 0.9836 | 0.9427 | 0.9373 |
| 0.9364 | 0.9647 | 0.9224 | 0.9099 |
",
+ "type": "table",
+ "image_path": "89ef14c3de00dd1d762baf07ab626bac7c2f3f10836603e833ad767ad6532fab.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 180,
+ 114,
+ 430,
+ 133.66666666666666
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 180,
+ 133.66666666666666,
+ 430,
+ 153.33333333333331
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 180,
+ 153.33333333333331,
+ 430,
+ 172.99999999999997
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 218,
+ 505,
+ 252
+ ],
+ "lines": [],
+ "index": 6,
+ "bbox_fs": [
+ 105,
+ 218,
+ 506,
+ 253
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 424
+ ],
+ "score": 0.967,
+ "type": "image",
+ "image_path": "b589cd81bc68f7c0374d34eb5d18286cf197a9628be19ee2e0751e9faaaf5554.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 183,
+ 273,
+ 425,
+ 286.72727272727275
+ ],
+ "spans": [],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 183,
+ 286.72727272727275,
+ 425,
+ 300.4545454545455
+ ],
+ "spans": [],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 183,
+ 300.4545454545455,
+ 425,
+ 314.18181818181824
+ ],
+ "spans": [],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 183,
+ 314.18181818181824,
+ 425,
+ 327.909090909091
+ ],
+ "spans": [],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 183,
+ 327.909090909091,
+ 425,
+ 341.63636363636374
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 183,
+ 341.63636363636374,
+ 425,
+ 355.3636363636365
+ ],
+ "spans": [],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 183,
+ 355.3636363636365,
+ 425,
+ 369.09090909090924
+ ],
+ "spans": [],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 183,
+ 369.09090909090924,
+ 425,
+ 382.818181818182
+ ],
+ "spans": [],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 183,
+ 382.818181818182,
+ 425,
+ 396.54545454545473
+ ],
+ "spans": [],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 183,
+ 396.54545454545473,
+ 425,
+ 410.2727272727275
+ ],
+ "spans": [],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 183,
+ 410.2727272727275,
+ 425,
+ 424.0000000000002
+ ],
+ "spans": [],
+ "index": 18
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 496
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 505,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 441,
+ 371,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Effects of dense re-feeding. Y-axis: evaluation RMSE,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 371,
+ 441,
+ 380,
+ 451
+ ],
+ "score": 0.28,
+ "content": "\\mathbf { X }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 380,
+ 441,
+ 505,
+ 453
+ ],
+ "score": 1.0,
+ "content": "-axis: epoch number. Baseline",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 505,
+ 465
+ ],
+ "score": 1.0,
+ "content": "model was trained with learning rate of 0.001. Applying re-feeding step with the same learning",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 476
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 506,
+ 476
+ ],
+ "score": 1.0,
+ "content": "rate almost did not help (Baseline RF). Learning rate of 0.005 (Baseline LR 0.005) is too big for",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 505,
+ 488
+ ],
+ "score": 1.0,
+ "content": "baseline model without re-feeding. However, increasing both learning rate and applying re-feeding",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 279,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 279,
+ 498
+ ],
+ "score": 1.0,
+ "content": "step clearly helps (Baseline LR 0.005 RF).",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 523,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "Applying dense re-feeding and increasing the learning rate, allowed us to further improve the evalua-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 505,
+ 547
+ ],
+ "score": 1.0,
+ "content": "tion RMSE from 0.9167 to 0.9100. Picking a checkpoint with best evaluation RMSE and computing",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 545,
+ 459,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 545,
+ 459,
+ 558
+ ],
+ "score": 1.0,
+ "content": "test RMSE gives as 0.9099, which we believe is significantly better than other methods.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 523,
+ 505,
+ 558
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 580,
+ 290,
+ 591
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 291,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 291,
+ 592
+ ],
+ "score": 1.0,
+ "content": "3.7 COMPARISON WITH OTHER METHODS",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 682
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 605,
+ 505,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We compare our best model with Recurrent Recommender Network from Wu et al. (2017) which",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "has been shown to outperform PMF (Mnih & Salakhutdinov, 2008), T-SVD (Koren, 2010) and I/U-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 639
+ ],
+ "score": 1.0,
+ "content": "AR (Sedhain et al., 2015) on the data we use (see Table 1 for data description). Note, that unlike",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 636,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "T-SVD and RRN, our method does not explicitly take into account temporal dynamics of ratings.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Yet, Table 3 shows that it is still capable of outperforming these methods on future rating prediction",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "task. We train each model using only the training set and compute evaluation RMSE for 100 epochs.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 670,
+ 424,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 424,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Then the checkpoint with the highest evaluation RMSE is tested on the test set.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 31,
+ "bbox_fs": [
+ 105,
+ 605,
+ 505,
+ 683
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "“Netflix 3 months” has 7 times less training data compared to “Netflix full”, it is therefore, not",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "surprising that the model’s performance is significantly worse if trained on this data alone (0.9373",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 709,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "vs 0.9099). In fact, the model that performs best on “Netflix full” over-fits on this set, and we had",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 390,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 390,
+ 732
+ ],
+ "score": 1.0,
+ "content": "to reduce the model’s complexity accordingly (see Table 4 for details).",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 687,
+ 506,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 112
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 100
+ ],
+ "score": 1.0,
+ "content": "Table 4: Test RMSE achieved by DeepRec on different Netflix subsets. All models are trained with",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 101,
+ 321,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 101,
+ 321,
+ 112
+ ],
+ "score": 1.0,
+ "content": "one iterative output re-feeding step per each iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "score": 0.979,
+ "html": "| DataSet | RMSE | Model Architecture |
| | |
| Netflix 3 months | 0.9373 | n,128,256,256,dp(0.65),256,128,n |
| Netflix 6 months | 0.9207 | n,256,256,512,dp(0.8),256,256,n |
| Netflix 1 year | 0.9225 | n,256,256,512,dp(0.8),256,256, n |
| Netfix Full | 0.9099 | n,512,512,1024,dp(0.8),512,512, n |
",
+ "type": "table",
+ "image_path": "89907b1ebad481827acd0332cbbaef5f201eabc2a54538f92e01b86f1cd34677.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 142.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 161,
+ 142.0,
+ 449,
+ 169.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 161,
+ 169.0,
+ 449,
+ 196.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 230,
+ 196,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 229,
+ 197,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 229,
+ 197,
+ 246
+ ],
+ "score": 1.0,
+ "content": "4 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 255,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Deep learning has revolutionized many areas of machine learning, and it is poised do so with recom-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 279
+ ],
+ "score": 1.0,
+ "content": "mender systems as well. In this paper we demonstrated how very deep autoencoders can be success-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "fully trained even on relatively small amounts of data by using both well established (dropout) and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "relatively recent (“scaled exponential linear units”) deep learning techniques. Further, we introduced",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "iterative output re-feeding - a technique which allowed us to perform dense updates in collaborative",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "filtering, increase learning rate and further improve generalization performance of our model. On",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "the task of future rating prediction, our model outperforms other approaches even without using",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 219,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 219,
+ 345
+ ],
+ "score": 1.0,
+ "content": "additional temporal signals.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "While our code supports item-based model (such as I-AutoRec) we argue that this approach is less",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "practical than user-based model (U-AutoRec). This is because in real-world recommender systems,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 383
+ ],
+ "score": 1.0,
+ "content": "there are usually much more users then items. Finally, when building personalized recommender",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 472,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 472,
+ 395
+ ],
+ "score": 1.0,
+ "content": "system and faced with scaling problems, it can be acceptable to sample items but not users.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 407,
+ 200,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 200,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 200,
+ 418
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 425,
+ 504,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "We thank the author of (Wu et al., 2017), Chao-Yuan Wu, for fruitfull discussion and help validating",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 161,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 161,
+ 448
+ ],
+ "score": 1.0,
+ "content": "our data sets.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 464,
+ 175,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 176,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 176,
+ 478
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 503,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 504,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 504,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Gediminas Adomavicius and Alexander Tuzhilin. Context-aware recommender systems. In Recom-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 116,
+ 494,
+ 344,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 494,
+ 344,
+ 506
+ ],
+ "score": 1.0,
+ "content": "mender systems handbook, pp. 217–253. Springer, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "James Bennett, Stan Lanning, and Netflix Netflix. The netflix prize. In In KDD Cup and Workshop",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 525,
+ 247,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 525,
+ 247,
+ 537
+ ],
+ "score": 1.0,
+ "content": "in conjunction with KDD, 2007.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "John S. Breese, David Heckerman, and Carl Kadie. Empirical analysis of predictive algorithms for",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 117,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "collaborative filtering. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 567,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 567,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Intelligence, UAI’98, pp. 43–52, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 577,
+ 504,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 577,
+ 504,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Inc. ISBN 1-55860-555-X. URL http://dl.acm.org/citation.cfm?id=2074094.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 588,
+ 163,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 588,
+ 163,
+ 599
+ ],
+ "score": 1.0,
+ "content": "2074100.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network´",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 451,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 451,
+ 631
+ ],
+ "score": 1.0,
+ "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 637,
+ 503,
+ 672
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. Neural col-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "laborative filtering. In Proceedings of the 26th International Conference on World Wide Web, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 659,
+ 444,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 659,
+ 444,
+ 672
+ ],
+ "score": 1.0,
+ "content": "173–182. International World Wide Web Conferences Steering Committee, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 679,
+ 503,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 116,
+ 690,
+ 302,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 690,
+ 302,
+ 702
+ ],
+ "score": 1.0,
+ "content": "networks. science, 313(5786):504–507, 2006.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 709,
+ 502,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 708,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 319,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Geoffrey E Hinton and Richard S. Zemel.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 708,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Autoencoders, minimum description",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 504,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 286,
+ 734
+ ],
+ "score": 1.0,
+ "content": "length and helmholtz free energy.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 720,
+ 504,
+ 732
+ ],
+ "score": 1.0,
+ "content": "In J. D. Cowan, G. Tesauro, and J. Al-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "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": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 112
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 504,
+ 100
+ ],
+ "score": 1.0,
+ "content": "Table 4: Test RMSE achieved by DeepRec on different Netflix subsets. All models are trained with",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 101,
+ 321,
+ 112
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 101,
+ 321,
+ 112
+ ],
+ "score": 1.0,
+ "content": "one iterative output re-feeding step per each iteration.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 196
+ ],
+ "score": 0.979,
+ "html": "| DataSet | RMSE | Model Architecture |
| | |
| Netflix 3 months | 0.9373 | n,128,256,256,dp(0.65),256,128,n |
| Netflix 6 months | 0.9207 | n,256,256,512,dp(0.8),256,256,n |
| Netflix 1 year | 0.9225 | n,256,256,512,dp(0.8),256,256, n |
| Netfix Full | 0.9099 | n,512,512,1024,dp(0.8),512,512, n |
",
+ "type": "table",
+ "image_path": "89907b1ebad481827acd0332cbbaef5f201eabc2a54538f92e01b86f1cd34677.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 161,
+ 115,
+ 449,
+ 142.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 161,
+ 142.0,
+ 449,
+ 169.0
+ ],
+ "spans": [],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 161,
+ 169.0,
+ 449,
+ 196.0
+ ],
+ "spans": [],
+ "index": 4
+ }
+ ]
+ }
+ ],
+ "index": 1.75
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 230,
+ 196,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 229,
+ 197,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 229,
+ 197,
+ 246
+ ],
+ "score": 1.0,
+ "content": "4 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 255,
+ 505,
+ 344
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 255,
+ 505,
+ 269
+ ],
+ "score": 1.0,
+ "content": "Deep learning has revolutionized many areas of machine learning, and it is poised do so with recom-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 504,
+ 279
+ ],
+ "score": 1.0,
+ "content": "mender systems as well. In this paper we demonstrated how very deep autoencoders can be success-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 278,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "fully trained even on relatively small amounts of data by using both well established (dropout) and",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 301
+ ],
+ "score": 1.0,
+ "content": "relatively recent (“scaled exponential linear units”) deep learning techniques. Further, we introduced",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "iterative output re-feeding - a technique which allowed us to perform dense updates in collaborative",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "filtering, increase learning rate and further improve generalization performance of our model. On",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 335
+ ],
+ "score": 1.0,
+ "content": "the task of future rating prediction, our model outperforms other approaches even without using",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 219,
+ 345
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 219,
+ 345
+ ],
+ "score": 1.0,
+ "content": "additional temporal signals.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 105,
+ 255,
+ 506,
+ 345
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 349,
+ 505,
+ 394
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 349,
+ 505,
+ 362
+ ],
+ "score": 1.0,
+ "content": "While our code supports item-based model (such as I-AutoRec) we argue that this approach is less",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 360,
+ 506,
+ 374
+ ],
+ "score": 1.0,
+ "content": "practical than user-based model (U-AutoRec). This is because in real-world recommender systems,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 383
+ ],
+ "score": 1.0,
+ "content": "there are usually much more users then items. Finally, when building personalized recommender",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 472,
+ 395
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 472,
+ 395
+ ],
+ "score": 1.0,
+ "content": "system and faced with scaling problems, it can be acceptable to sample items but not users.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 349,
+ 506,
+ 395
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 407,
+ 200,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 200,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 407,
+ 200,
+ 418
+ ],
+ "score": 1.0,
+ "content": "ACKNOWLEDGMENTS",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 425,
+ 504,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 506,
+ 439
+ ],
+ "score": 1.0,
+ "content": "We thank the author of (Wu et al., 2017), Chao-Yuan Wu, for fruitfull discussion and help validating",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 436,
+ 161,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 436,
+ 161,
+ 448
+ ],
+ "score": 1.0,
+ "content": "our data sets.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 105,
+ 423,
+ 506,
+ 448
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 464,
+ 175,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 176,
+ 478
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 465,
+ 176,
+ 478
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 483,
+ 503,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 504,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 504,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Gediminas Adomavicius and Alexander Tuzhilin. Context-aware recommender systems. In Recom-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 116,
+ 494,
+ 344,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 494,
+ 344,
+ 506
+ ],
+ "score": 1.0,
+ "content": "mender systems handbook, pp. 217–253. Springer, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 482,
+ 504,
+ 506
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 513,
+ 504,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "James Bennett, Stan Lanning, and Netflix Netflix. The netflix prize. In In KDD Cup and Workshop",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 525,
+ 247,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 525,
+ 247,
+ 537
+ ],
+ "score": 1.0,
+ "content": "in conjunction with KDD, 2007.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 511,
+ 505,
+ 537
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 543,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 544,
+ 505,
+ 556
+ ],
+ "score": 1.0,
+ "content": "John S. Breese, David Heckerman, and Carl Kadie. Empirical analysis of predictive algorithms for",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 117,
+ 556,
+ 505,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 556,
+ 505,
+ 568
+ ],
+ "score": 1.0,
+ "content": "collaborative filtering. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 567,
+ 505,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 567,
+ 505,
+ 578
+ ],
+ "score": 1.0,
+ "content": "Intelligence, UAI’98, pp. 43–52, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 577,
+ 504,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 577,
+ 504,
+ 590
+ ],
+ "score": 1.0,
+ "content": "Inc. ISBN 1-55860-555-X. URL http://dl.acm.org/citation.cfm?id=2074094.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 115,
+ 588,
+ 163,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 588,
+ 163,
+ 599
+ ],
+ "score": 1.0,
+ "content": "2074100.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 28,
+ "bbox_fs": [
+ 106,
+ 544,
+ 505,
+ 599
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 606,
+ 504,
+ 630
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "Djork-Arne Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network´",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 451,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 451,
+ 631
+ ],
+ "score": 1.0,
+ "content": "learning by exponential linear units (elus). arXiv preprint arXiv:1511.07289, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 31.5,
+ "bbox_fs": [
+ 106,
+ 606,
+ 505,
+ 631
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 637,
+ 503,
+ 672
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 637,
+ 505,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. Neural col-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "laborative filtering. In Proceedings of the 26th International Conference on World Wide Web, pp.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 659,
+ 444,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 659,
+ 444,
+ 672
+ ],
+ "score": 1.0,
+ "content": "173–182. International World Wide Web Conferences Steering Committee, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 106,
+ 637,
+ 505,
+ 672
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 679,
+ 503,
+ 702
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 505,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 679,
+ 505,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 116,
+ 690,
+ 302,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 690,
+ 302,
+ 702
+ ],
+ "score": 1.0,
+ "content": "networks. science, 313(5786):504–507, 2006.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 106,
+ 679,
+ 505,
+ 702
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 709,
+ 502,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 708,
+ 505,
+ 721
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 709,
+ 319,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Geoffrey E Hinton and Richard S. Zemel.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 708,
+ 505,
+ 721
+ ],
+ "score": 1.0,
+ "content": "Autoencoders, minimum description",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 718,
+ 504,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 718,
+ 286,
+ 734
+ ],
+ "score": 1.0,
+ "content": "length and helmholtz free energy.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 295,
+ 720,
+ 504,
+ 732
+ ],
+ "score": 1.0,
+ "content": "In J. D. Cowan, G. Tesauro, and J. Al-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "spector (eds.), Advances in Neural Information Processing Systems 6, pp. 3–",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 92,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 92,
+ 231,
+ 106
+ ],
+ "score": 1.0,
+ "content": "10. Morgan-Kaufmann,",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 236,
+ 93,
+ 264,
+ 105
+ ],
+ "score": 1.0,
+ "content": "1994.",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 299,
+ 92,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "URL http://papers.nips.cc/paper/",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 103,
+ 541,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 103,
+ 541,
+ 118
+ ],
+ "score": 1.0,
+ "content": "798-autoencoders-minimum-description-length-and-helmholtz-free-energy.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 114,
+ 116,
+ 139,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 116,
+ 139,
+ 126
+ ],
+ "score": 1.0,
+ "content": "pdf.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 107,
+ 708,
+ 505,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 115,
+ 82,
+ 538,
+ 127
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 115,
+ 82,
+ 505,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 82,
+ 505,
+ 95
+ ],
+ "score": 1.0,
+ "content": "spector (eds.), Advances in Neural Information Processing Systems 6, pp. 3–",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 92,
+ 505,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 92,
+ 231,
+ 106
+ ],
+ "score": 1.0,
+ "content": "10. Morgan-Kaufmann,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 93,
+ 264,
+ 105
+ ],
+ "score": 1.0,
+ "content": "1994.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 299,
+ 92,
+ 505,
+ 106
+ ],
+ "score": 1.0,
+ "content": "URL http://papers.nips.cc/paper/",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 103,
+ 541,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 103,
+ 541,
+ 118
+ ],
+ "score": 1.0,
+ "content": "798-autoencoders-minimum-description-length-and-helmholtz-free-energy.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 114,
+ 116,
+ 139,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 116,
+ 139,
+ 126
+ ],
+ "score": 1.0,
+ "content": "pdf.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 1.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 134,
+ 505,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Hyunsoo Kim and Haesun Park. Nonnegative matrix factorization based on alternating nonneg-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 117,
+ 146,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 146,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "ativity constrained least squares and active set method. SIAM journal on matrix analysis and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 156,
+ 259,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 156,
+ 259,
+ 168
+ ],
+ "score": 1.0,
+ "content": "applications, 30(2):713–730, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 174,
+ 504,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Gunter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing ¨",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 186,
+ 348,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 186,
+ 348,
+ 198
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1706.02515, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 504,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren. The bellkor solution to the netflix grand prize. Netflix prize documentation, 81:1–10,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 215,
+ 142,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 215,
+ 142,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2009.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 234,
+ 504,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren. Collaborative filtering with temporal dynamics. Communications of the ACM, 53",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 246,
+ 186,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 246,
+ 186,
+ 258
+ ],
+ "score": 1.0,
+ "content": "(4):89–97, 2010.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 264,
+ 504,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 276,
+ 250,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 276,
+ 250,
+ 288
+ ],
+ "score": 1.0,
+ "content": "systems. Computer, 42(8), 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 294,
+ 504,
+ 318
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 114,
+ 304,
+ 144,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 304,
+ 144,
+ 318
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 324,
+ 504,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Sheng Li, Jaya Kawale, and Yun Fu. Deep collaborative filtering via marginalized denoising auto-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "encoder. In Proceedings of the 24th ACM International on Conference on Information and Knowl-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 346,
+ 301,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 346,
+ 301,
+ 358
+ ],
+ "score": 1.0,
+ "content": "edge Management, pp. 811–820. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 504,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Andriy Mnih and Ruslan R Salakhutdinov. Probabilistic matrix factorization. In Advances in neural",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 376,
+ 336,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 376,
+ 336,
+ 389
+ ],
+ "score": 1.0,
+ "content": "information processing systems, pp. 1257–1264, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 394,
+ 503,
+ 418
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 114,
+ 406,
+ 254,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 406,
+ 254,
+ 417
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1710.05941, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Francesco Ricci, Lior Rokach, and Bracha Shapira. Introduction to recommender systems hand-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 436,
+ 393,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 436,
+ 393,
+ 448
+ ],
+ "score": 1.0,
+ "content": "book. In Recommender systems handbook, pp. 1–35. Springer, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 504,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton. Restricted boltzmann machines for",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 465,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 465,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "collaborative filtering. In Proceedings of the 24th international conference on Machine learning,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 477,
+ 224,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 477,
+ 224,
+ 488
+ ],
+ "score": 1.0,
+ "content": "pp. 791–798. ACM, 2007.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 495,
+ 504,
+ 529
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "Suvash Sedhain, Aditya Krishna Menon, Scott Sanner, and Lexing Xie. Autorec: Autoencoders",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 507,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 507,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "meet collaborative filtering. In Proceedings of the 24th International Conference on World Wide",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 117,
+ 517,
+ 245,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 517,
+ 245,
+ 531
+ ],
+ "score": 1.0,
+ "content": "Web, pp. 111–112. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 536,
+ 504,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Florian Strub and Jer´ emie Mary. Collaborative filtering with stacked denoising autoencoders and ´",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 426,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 426,
+ 560
+ ],
+ "score": 1.0,
+ "content": "sparse inputs. In NIPS workshop on machine learning for eCommerce, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 566,
+ 504,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 504,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 504,
+ 580
+ ],
+ "score": 1.0,
+ "content": "Hao Wang, Naiyan Wang, and Dit-Yan Yeung. Collaborative deep learning for recommender sys-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "tems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discov-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 589,
+ 322,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 589,
+ 322,
+ 600
+ ],
+ "score": 1.0,
+ "content": "ery and Data Mining, pp. 1235–1244. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 607,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J. Smola, and How Jing. Recurrent recom-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 117,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "mender networks. In Proceedings of the Tenth ACM International Conference on Web Search",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 629,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 629,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "and Data Mining, WSDM ’17, pp. 495–503, New York, NY, USA, 2017. ACM. ISBN 978-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "1-4503-4675-7. doi: 10.1145/3018661.3018689. URL http://doi.acm.org/10.1145/",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 651,
+ 212,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 651,
+ 212,
+ 663
+ ],
+ "score": 1.0,
+ "content": "3018661.3018689.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 669,
+ 503,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li. Empirical evaluation of rectified activations in",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 681,
+ 374,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 681,
+ 374,
+ 693
+ ],
+ "score": 1.0,
+ "content": "convolutional network. arXiv preprint arXiv:1505.00853, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 700,
+ 504,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Yin Zheng, Bangsheng Tang, Wenkui Ding, and Hanning Zhou. A neural autoregressive approach",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 381,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 381,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to collaborative filtering. arXiv preprint arXiv:1605.09477, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 308,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 26,
+ 309,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Under review as a conference paper at ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 115,
+ 82,
+ 538,
+ 127
+ ],
+ "lines": [],
+ "index": 1.5,
+ "bbox_fs": [
+ 114,
+ 82,
+ 541,
+ 126
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 134,
+ 505,
+ 168
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 149
+ ],
+ "score": 1.0,
+ "content": "Hyunsoo Kim and Haesun Park. Nonnegative matrix factorization based on alternating nonneg-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 117,
+ 146,
+ 505,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 146,
+ 505,
+ 157
+ ],
+ "score": 1.0,
+ "content": "ativity constrained least squares and active set method. SIAM journal on matrix analysis and",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 156,
+ 259,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 156,
+ 259,
+ 168
+ ],
+ "score": 1.0,
+ "content": "applications, 30(2):713–730, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 132,
+ 505,
+ 168
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 174,
+ 504,
+ 198
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 189
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 173,
+ 506,
+ 189
+ ],
+ "score": 1.0,
+ "content": "Gunter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing ¨",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 186,
+ 348,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 186,
+ 348,
+ 198
+ ],
+ "score": 1.0,
+ "content": "neural networks. arXiv preprint arXiv:1706.02515, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 173,
+ 506,
+ 198
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 204,
+ 504,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 204,
+ 505,
+ 217
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren. The bellkor solution to the netflix grand prize. Netflix prize documentation, 81:1–10,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 215,
+ 142,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 215,
+ 142,
+ 228
+ ],
+ "score": 1.0,
+ "content": "2009.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 106,
+ 204,
+ 505,
+ 228
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 234,
+ 504,
+ 258
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 234,
+ 505,
+ 248
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren. Collaborative filtering with temporal dynamics. Communications of the ACM, 53",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 246,
+ 186,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 246,
+ 186,
+ 258
+ ],
+ "score": 1.0,
+ "content": "(4):89–97, 2010.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 106,
+ 234,
+ 505,
+ 258
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 264,
+ 504,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 264,
+ 505,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Yehuda Koren, Robert Bell, and Chris Volinsky. Matrix factorization techniques for recommender",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 276,
+ 250,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 276,
+ 250,
+ 288
+ ],
+ "score": 1.0,
+ "content": "systems. Computer, 42(8), 2009.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 264,
+ 505,
+ 288
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 294,
+ 504,
+ 318
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 293,
+ 505,
+ 308
+ ],
+ "score": 1.0,
+ "content": "Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 114,
+ 304,
+ 144,
+ 318
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 304,
+ 144,
+ 318
+ ],
+ "score": 1.0,
+ "content": "2015.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 105,
+ 293,
+ 505,
+ 318
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 324,
+ 504,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 504,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Sheng Li, Jaya Kawale, and Yun Fu. Deep collaborative filtering via marginalized denoising auto-",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 116,
+ 336,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 336,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "encoder. In Proceedings of the 24th ACM International on Conference on Information and Knowl-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 346,
+ 301,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 346,
+ 301,
+ 358
+ ],
+ "score": 1.0,
+ "content": "edge Management, pp. 811–820. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 105,
+ 324,
+ 505,
+ 358
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 365,
+ 504,
+ 388
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 505,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Andriy Mnih and Ruslan R Salakhutdinov. Probabilistic matrix factorization. In Advances in neural",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 376,
+ 336,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 376,
+ 336,
+ 389
+ ],
+ "score": 1.0,
+ "content": "information processing systems, pp. 1257–1264, 2008.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 106,
+ 364,
+ 505,
+ 389
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 394,
+ 503,
+ 418
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 394,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "Prajit Ramachandran, Barret Zoph, and Quoc V Le. Swish: a self-gated activation function. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 114,
+ 406,
+ 254,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 406,
+ 254,
+ 417
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1710.05941, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 106,
+ 394,
+ 505,
+ 417
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 424,
+ 504,
+ 448
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 505,
+ 438
+ ],
+ "score": 1.0,
+ "content": "Francesco Ricci, Lior Rokach, and Bracha Shapira. Introduction to recommender systems hand-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 436,
+ 393,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 436,
+ 393,
+ 448
+ ],
+ "score": 1.0,
+ "content": "book. In Recommender systems handbook, pp. 1–35. Springer, 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24.5,
+ "bbox_fs": [
+ 105,
+ 424,
+ 505,
+ 448
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 454,
+ 504,
+ 489
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 455,
+ 505,
+ 467
+ ],
+ "score": 1.0,
+ "content": "Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton. Restricted boltzmann machines for",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 465,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 465,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "collaborative filtering. In Proceedings of the 24th international conference on Machine learning,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 477,
+ 224,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 477,
+ 224,
+ 488
+ ],
+ "score": 1.0,
+ "content": "pp. 791–798. ACM, 2007.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 106,
+ 455,
+ 506,
+ 488
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 495,
+ 504,
+ 529
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 495,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "Suvash Sedhain, Aditya Krishna Menon, Scott Sanner, and Lexing Xie. Autorec: Autoencoders",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 507,
+ 505,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 507,
+ 505,
+ 519
+ ],
+ "score": 1.0,
+ "content": "meet collaborative filtering. In Proceedings of the 24th International Conference on World Wide",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 117,
+ 517,
+ 245,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 517,
+ 245,
+ 531
+ ],
+ "score": 1.0,
+ "content": "Web, pp. 111–112. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 106,
+ 495,
+ 505,
+ 531
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 536,
+ 504,
+ 560
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 505,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Florian Strub and Jer´ emie Mary. Collaborative filtering with stacked denoising autoencoders and ´",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 426,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 426,
+ 560
+ ],
+ "score": 1.0,
+ "content": "sparse inputs. In NIPS workshop on machine learning for eCommerce, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 536,
+ 505,
+ 560
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 566,
+ 504,
+ 601
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 504,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 504,
+ 580
+ ],
+ "score": 1.0,
+ "content": "Hao Wang, Naiyan Wang, and Dit-Yan Yeung. Collaborative deep learning for recommender sys-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 116,
+ 577,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 577,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "tems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discov-",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 589,
+ 322,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 589,
+ 322,
+ 600
+ ],
+ "score": 1.0,
+ "content": "ery and Data Mining, pp. 1235–1244. ACM, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 105,
+ 565,
+ 505,
+ 600
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 607,
+ 505,
+ 663
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Chao-Yuan Wu, Amr Ahmed, Alex Beutel, Alexander J. Smola, and How Jing. Recurrent recom-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 117,
+ 619,
+ 505,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 619,
+ 505,
+ 631
+ ],
+ "score": 1.0,
+ "content": "mender networks. In Proceedings of the Tenth ACM International Conference on Web Search",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 629,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 629,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "and Data Mining, WSDM ’17, pp. 495–503, New York, NY, USA, 2017. ACM. ISBN 978-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "1-4503-4675-7. doi: 10.1145/3018661.3018689. URL http://doi.acm.org/10.1145/",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 116,
+ 651,
+ 212,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 651,
+ 212,
+ 663
+ ],
+ "score": 1.0,
+ "content": "3018661.3018689.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 39,
+ "bbox_fs": [
+ 106,
+ 606,
+ 505,
+ 663
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 669,
+ 503,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 505,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li. Empirical evaluation of rectified activations in",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 681,
+ 374,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 681,
+ 374,
+ 693
+ ],
+ "score": 1.0,
+ "content": "convolutional network. arXiv preprint arXiv:1505.00853, 2015.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 106,
+ 670,
+ 505,
+ 693
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 700,
+ 504,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "Yin Zheng, Bangsheng Tang, Wenkui Ding, and Hanning Zhou. A neural autoregressive approach",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 381,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 381,
+ 723
+ ],
+ "score": 1.0,
+ "content": "to collaborative filtering. arXiv preprint arXiv:1605.09477, 2016.",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 106,
+ 699,
+ 505,
+ 723
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file