reports/paper_writing_guide.md

CS3319 论文写作辅助文档(Paper Writing Guide)

本文档为论文撰写的"素材库 + 脚手架"。所有数字均来自项目真实实验产物(各 validation_summary.csv / *_ablation.csv / error_analysis_buckets.csv / dynamic_summary.csv / notes/experiment_history.md / 图结构统计),可直接核对、直接复制进论文。验证集 F1 均基于同一 split_seed=202 划分,严格可比。

---

0. 一页速览(TL;DR)

项	内容
任务	学术异构图上的作者-论文链接预测(reading recommendation),F1 评估
数据	6,611 作者 / 79,937 论文;合著边 9,663 / 引用边 327,113 / 训练阅读边 682,421 / 测试对 2,047,262;训练图密度仅 0.129%
最终方法	两阶段堆叠:Stage1 多个嵌入/分数源(LightGCN、BPR-MF、DeepWalk、Node2Vec)→ 259 维特征工程 → Stage2 LightGBM 二级学习器 → rank-cutoff 决策
最终成绩	验证 F1 = 0.9669,AUC = 0.9949;公开 LB F1 = 0.9663
核心叙事	方法演进"三级跳":0.939(单 GNN)→ 0.956(特征堆叠)→ 0.967(高阶引用传播)
三大卖点	① 高阶有向引用传播(A-P-P^k、A-A-P-P^k,fwd/bwd/undir);② 多源异构特征 + 二级 LightGBM 堆叠;③ rank-cutoff 决策规避验证-测试分布漂移

---

1. 论文定位与核心卖点

1.1 定位

这是一个课程项目向正式论文靠拢的工作(参考 data_and_docs/advice.md 的 reviewer 反馈)。建议定位为 "基于异构图特征工程与高阶传播的学术阅读推荐",强调:

• 不是"盲目堆复杂模型",而是根据任务特点(稀疏 + 冷启动)选择 LightGCN + 显式高阶传播;
• 方法可解释、可复现、CPU 即可跑。

1.2 三个核心卖点(每个都要在论文里给证据)

卖点	证据	出处
高阶引用传播带来第三级突破	+high-order directed 使 F1 从 0.9650 → 0.9669(+0.0019),n_features 190→259	high_order_graph_stack/validation_summary.csv
特征堆叠远胜单一 GNN	LightGCN 单模型 0.9386 → 堆叠 0.9559(+0.0173,最大单步增益)	dynamic_summary.csv + post95_ablation/ablation_table.csv
rank-cutoff 决策更稳健	验证集 1:1 人工划分导致概率阈值在测试上漂移到 0.52;rank-cutoff 在 ratio∈[0.498,0.502] 上 F1 几乎不变	stack_ratio_analysis.csv + threshold_submission_summary.csv

1.3 一句话主旨(可放 Introduction 末尾)

We show that, on a sparse academic reading-recommendation graph, higher-order directed citation propagation combined with a multi-source feature stacker yields large gains over single-graph neural models, and that a rank-based decision rule is markedly more robust than probability thresholding under the validation–test distribution shift.

---

2. 建议的论文结构与逐章写作指南

Abstract(模板,中英可填)

Link prediction on heterogeneous academic graphs underpins reading-recommendation systems, yet suffers from severe sparsity and cold start. We study the author–paper recommendation task on a GeoScience academic network (6,611 authors, 79,937 papers). We propose a two-stage framework: (1) multiple embedding/score producers — LightGCN, BPR-MF, DeepWalk, and Node2Vec — and (2) a LightGBM second-stage stacker over 259 engineered features, including higher-order directed citation propagation along Author-Paper-Paper^k and Author-Author-Paper-Paper^k meta-paths. On a held-out validation split our method reaches F1 = 0.9669 (AUC = 0.9949), improving over the best single LightGCN (0.9386) by +2.8 points, and over the content/feature stacker (0.9559) by +1.1 points; the public-leaderboard F1 is 0.9663. We further show that a rank-cutoff decision rule is more stable than probability thresholding under distribution shift.

1 Introduction

要讲清四件事(呼应 advice.md):

1. 问题定义清晰:author-paper link prediction,不是普通分类——这是 reviewer 最看重的点。给出图怎么建、节点是什么、边是什么、label 是什么、prediction target 是什么。
2. 冷启动是 GNN 在推荐中的核心价值:图结构能传播稀疏节点信息(56.4% 的作者只有 1 个合著者;12,042 篇论文从未被读、16,074 篇从未被引)。
3. 动机:单 GNN 受限于稀疏与未校准概率;我们用高阶传播 + 堆叠 + rank 决策三管齐下。
4. 贡献(contributions,用 bullet):
   • 设计有向/无向高阶引用传播特征族(24+45 维);
   • 系统集成 7 个随机游走配置 + LightGCN/BPR-MF/内容特征 为 259 维二级特征;
   • 揭示验证-测试分布漂移,提出 rank-cutoff 决策;
   • 充分的消融、误差分析与跨种子稳定性。

2 Related Work

建议三层 baseline 框架(直接采用 advice.md 的建议):

• Level 1 启发式:popularity、common neighbor、Adamic-Adar(在 author-paper 二部图上);
• Level 2 传统图嵌入:BPR / implicit MF、DeepWalk、Node2Vec;
• Level 3 GNN:LightGCN、(Hetero)GNN。 本工作 = Level 2/3 多源融合 + 高阶 meta-path 特征,位于 Level 3 之上。需补一句 meta-path 经典性:Author→Author→Paper、Author→Paper→Paper 是异构图推荐的关键词。

3 Problem Formulation & Dataset

• 形式化:异构图 $G=(V,E)$,$V=A\cup P$(作者、论文),边类型 ${$reads, co-author, cites$}$。给定测试对集合 $\mathcal{T}\subseteq A\times P$,预测 $\hat y\in{0,1}$。
• **数据集统计表(见 §3.1,可直接放论文 Table 1)**。
• 评估:F1;公开榜基于 50% 测试集。
• ⚠️ 必须写一句防 label leakage(advice.md 强调):

  "We carefully avoid information leakage: citation/co-authorship edges are built only from the provided static networks, and validation edges are held out of the LightGCN message-passing graph."

4 Method(论文核心章)

按四小节组织,每节给可复述的素材:

(a) Stage-1 Score Producers

• LightGCN:异构 4 层,边类型 (author,ref,paper)/(paper,beref,author)/(paper,cite,paper)/(author,coauthor,author);作者用可训练 embedding,论文投影 feature.pkl(512 维 USE)→ embed;层输出加权求和;BPR 损失 + 硬负采样(random 50% / popular 25% / co-author pool 25%)。最佳配置 L=2, d=512, BPR。
• BPR-MF:Embedding(6611,256)×Embedding(79937,256)+bias,AdamW,220 epoch。
• DeepWalk / Node2Vec:7 个配置(见 §4.3),每个生成 11 维 pair 特征。

(b) Feature Engineering(259 维,见 §4.2 完整构成表)

• 显式图/meta-path 特征(18 维:aap/app/apap 计数与比率、Jaccard、度等);
• 高阶引用传播(见下,创新点);
• content-rich(18 维)、content mean-cos、top-k 相似度、负证据、变体分数。

(c) Higher-Order Directed Citation Propagation(创新点,单独一小节)

• 在行归一化邻接矩阵上做稀疏传播 + top-k 剪枝(k=1500):
  • A-P-P^k(k=1..4):作者经由阅读论文沿引用图传播;
  • A-A-P-P^k(k=0..3):先经合著聚合再传播;
  • 三方向:forward / backward / undirected 引用;
• 每跳产出 raw / popularity-normalized(除以 log1p(paper_ref_deg+cite_deg))/ log / delta 变体。
• 论证:冷启动作者(阅读历史少)可通过合著者与引用社区获得信号——这正是 GNN 在稀疏推荐中的价值。

(d) Stage-2 LightGBM Stacker & Decision Rule

• 5 折 StratifiedKFold 出 OOF 分数用于无泄漏评估;
• 最终超参:num_leaves=15, reg_lambda=8.0, min_child_samples=100, n_estimators=1400, lr=0.022(强正则,防过拟合);
• 决策规则(关键):按最终分数排序,预测 top-50% 为正,并强制训练-测试重叠的已知正例为 1(占测试集 25.6%)。

5 Experiments

• 5.1 Setup:split_seed=202,train_frac=0.9 的"notebook-style"划分(10% 训练边作验证正例 + 等量随机负例,1:1)。说明这是人工 1:1 划分,故阈值不直接迁移到测试。
• 5.2 Main Results / Progression(Table 2,见 §3.2)——最重要的表。
• 5.3 Ablation(Table 3,见 §3.3):逐步加入特征族的增益。
• 5.4 Architecture Sensitivity(Table 4 / 图):LightGCN dim/layers/loss。
• 5.5 Cross-seed Stability:F1 方差 < 0.001。

6 Analysis & Discussion

• 6.1 Error Analysis by Group(见 §3.4):has_local_evidence=0 的 F1 仅 0.573 vs =1 的 0.969——冷启动/无局部证据是主要误差源;按度分桶的低度节点性能下降。
• 6.2 Calibration & Decision Robustness:校准曲线 + ratio-F1 曲线,论证 rank-cutoff。
• 6.3 Feature Importance(若导出):高阶传播族贡献占比。
• 6.4 Limitations:数据匿名无领域标签;验证集人工 1:1 与真实正例比例不符;rank top-50% 的假设。

7 Conclusion

回扣三级跳 + 三个贡献 + 未来工作(可加领域标签/图注意力/校准方法如 Platt/Isotonic)。

---

3. 核心数据表格(可直接复制进论文)

3.1 Table 1 — Dataset Statistics

Entity / Relation	Count	Note
Authors	6,611	—
Papers	79,937	512-d USE features
Co-authorship edges	9,663	undirected
Citation edges	327,113	directed
Train read edges (author–paper)	682,421	positives
Test pairs	2,047,262	to predict
Bipartite density	1.29×10⁻³	0.129%
Co-author connected components	1,508	largest = 873
Authors with degree = 1	56.4%	severe sparsity
Papers never read	12,042	cold-start items
Papers never cited	16,074	—

度分布(用于 log-log 图):

• 合著度:median=1, P90=6, P99=28, max=88
• 论文被读度:median=4, mean=10.05, P99=103, max=2,582
• 论文被引(入)度:median=3, mean=5.12, P99=34, max=777
• 作者阅读数(训练):median=74, mean=103.4, max=5,727

3.2 Table 2 — Performance Progression(主角表 ★)

#	Method	Val F1	ΔF1	#Feat	Public LB F1
0	Official notebook baseline (HeteroMeanConv+hinge)	~0.885	—	—	—
1	LightGCN single (L2, d512, BPR, best)	0.9386	—	—	0.9304 (6-model ens)
2	+ explicit graph/meta-path + rank (stacking baseline)	0.9560	+0.0174	22	—
3	+ neg-evidence + top-k + variant scores	0.9571	+0.0011	76	0.9576
4	+ content mean-cos	0.9576	+0.0005	80	—
5	+ BPR-MF	0.9593	+0.0017	84	0.9600
6	+ rich content (18-d)	0.9599	+0.0006	102	—
7	+ DeepWalk + Node2Vec	0.9621	+0.0022	100	0.9625
8	+ 7-block random-walk ensemble	0.9649	+0.0028	190	—
9	+ high-order propagation (undirected)	0.9666	+0.0017	214	—
10	+ high-order directed (FINAL)	0.9669	+0.0003	259	0.9663

三级跳标注:① +0.0173(行 1→2,堆叠)② +0.0030(行 6→8,随机游走)③ +0.0020(行 8→10,高阶传播)。

3.3 Table 3 — Feature-Group Ablation(同 split,逐步增益)

Stage	Val F1	n_features	Source CSV
baseline_stacking	0.9560	22	post95_ablation/ablation_table.csv
+ ensemble_lgcn_score_features	0.9571	76	同上
+ content_mean_cos	0.9576	80	extra_score_sources/extra_score_ablation.csv
+ bpr_mf	0.9593	84	同上
+ rich_content_features	0.9599	102	content_rich/content_rich_ablation.csv
+ node2vec	0.9621	100	node2vec_deepwalk/node2vec_deepwalk_ablation.csv
+ deepwalk	0.9618	92	同上
rwens_7model	0.9649	172	randomwalk_systematic/ensemble_7_ablation.csv

3.4 Table 4 — High-Order Propagation Ablation(创新点消融 ★)

Stage	Val F1	AUC	Precision	Recall	n_features
base_highorder	0.9643	0.9941	0.9654	0.9632	108
rich_rw7	0.9650	0.9946	0.9664	0.9635	190
rich_rw7_highorder	0.9666	0.9949	0.9671	0.9660	214
rich_rw7_highorder_directed	0.9669	0.9949	0.9667	0.9670	259

3.5 Table 5 — Error Analysis by Group(支撑 Discussion)

Group	Bucket	F1	Insight
has_local_evidence	0	0.573	无局部证据 = 主要误差源
has_local_evidence	1	0.969	有局部证据几乎全对
author_degree	[-inf,1)	0.500	冷启动作者极差
author_degree	[1,3)	0.667	低度作者差
author_degree	[50,inf)	高(见 CSV)	高度作者好
LightGCN_score	[0.51,1.27]	0.275	中间分数段严重误判(校准差)

3.6 Table 6 — Decision Rule Robustness(支撑 rank-cutoff 论点)

Rule	Positive ratio on test	Stability
probability threshold th=0.455	0.5250	漂移到 52.5%
probability threshold th=0.500	0.5199	仍漂移
rank-cutoff ratio=0.500	0.5000	稳定 50%

ratio∈[0.498, 0.502] 时验证 F1 几乎不变(0.9554–0.9559,stacking 阶段;final 阶段全 0.9669)。

3.7 LightGCN 架构敏感性(支撑 5.4)

• 最佳:L=2, d=512, BPR → 0.9386;L=3,d=512 → 0.9381(接近)。
• 损失函数:BPR(0.9386)≫ BCE(0.8612);hinge 居中。BPR 显著优于 BCE 是一个干净结论。
• 跨 seed(201–204)F1 ∈ [0.9365, 0.9386],方差 < 0.001 → 稳定性强。

---

4. 方法描述素材(可直接改写进 Method 章)

4.1 LightGCN(Stage-1 主力)

LightGCN adapted to the heterogeneous academic graph. Author nodes use learnable embeddings; paper nodes project the 512-d USE features via a linear layer. We stack L propagation layers that aggregate over four edge types — (author, ref, paper), (paper, beref, author), (paper, cite, paper), (author, coauthor, author) — each followed by symmetric-degree normalization. The final representation is the uniform (or learnable-weighted) mean over all layers, following LightGCN. The model is trained with the BPR ranking loss and hard negative sampling: 50% random negatives, 25% sampled from popular papers (top-30% by read-degree), and 25% from each author's co-author paper pool — explicitly targeting confusable items in the same citation/co-author community.

4.2 259 维特征完整构成(放 Method 或附录)

组	维数	说明
LightGCN rank	4	score / global-rank / author-pct / author-rank
Explicit graph & meta-path	18	aap/app/apap 计数与比率、Jaccard(ref/cited-by)、度等
Negative evidence	8	反向证据特征
Top-k content similarity	3	top-1/3/5 USE cosine
LightGCN variant scores	43	20 个变体×(z,rank)+3 聚合
content mean-cos	4	raw/z/rank/author-rank
BPR-MF	4	raw/z/rank/author-rank
X_base 小计	84
Rich content	18	center-cos、top-k、frac>0.5/0.7、local-pct 等
7× RW blocks	77	每块 11:dot/cos/hadamard/absdiff/l2 + 全局/作者 rank & pct
RW aggregate	11	7 块的 mean/std/max/min + agreement
High-order undirected	24	A-P-P^k,A-A-P-P^k × {raw,popnorm,log}
High-order directed	45	fwd/bwd/undir × {raw,popnorm,delta}
总计	259

4.3 七个随机游走配置(放 Method/附录)

Version	Graph	Method	dim	walk_len
dw_base_d128_l40	full	DeepWalk	128	40
dw_long_d128_l80	full	DeepWalk	128	80
dw_highdim_d256_l40	full	DeepWalk	256	40
dw_d256_l80	full	DeepWalk	256	80
dw_seed3407_d128_l40	full	DeepWalk	128	40
dw_graph_ap_pp	AP+PP	DeepWalk	128	40
n2v_p2_q1_d128_l40	full	Node2Vec(p=2,q=1)	128	40

RW ensemble 规模实验:1→5→7 模型,F1 0.9627→0.9639→0.9649,多样性带来稳定增益。

4.4 高阶传播(创新点,精炼描述)

We construct higher-order citation-propagation features by iterative sparse matrix multiplication on row-normalized adjacency matrices with top-k row pruning (k=1500). Two meta-path families are used: A-P-P^k (an author's read papers propagated k hops along the citation graph, k=1..4) and A-A-P-P^k (first aggregated through co-authors, then propagated, k=0..3). To disentangle citation directionality, we build forward / backward / undirected variants. For each hop we emit raw, popularity-normalized, log-transformed, and inter-hop delta signals, yielding 24 (undirected) + 45 (directed) = 69 features. These features let cold-start authors inherit signal from their co-author and citation communities — the core value of graph structure in sparse recommendation.

4.5 决策规则(精炼描述)

Because the validation split is an artificial 1:1 positive/negative construction, the LightGBM probabilities are not calibrated to the test distribution: a probability threshold tuned on validation drifts the test positive rate to ≈0.52. We therefore adopt a rank-cutoff decision rule: sort test pairs by final score, predict the top-50% as positive, and force the 25.6% of test pairs that overlap the training graph to positive. This is markedly more stable — F1 is flat across ratio ∈ [0.498, 0.502].

---

5. 图表规划清单(每张图的位置/目的/caption)

图	章节	目的	Caption 模板
Fig 1 Dataset overview	§3	稀疏/冷启动	"Degree distributions (log-log) and graph statistics of the academic network; 56% of authors have a single co-author."
Fig 2 Framework	§4	方法总览	"Two-stage stacking framework: score producers → 259-d feature engineering → LightGBM → rank-cutoff decision."
Fig 3 Progression ★	§5.2	三级跳	"Validation F1 progression. Three breakthroughs: +0.017 (stacking), +0.003 (random walks), +0.002 (higher-order propagation)."
Fig 4 High-order ablation	§5.3	创新点	"Ablation of higher-order directed citation propagation; adding directed propagation yields the best F1 (0.9669) at 259 features."
Fig 5 PR/ROC	§5.2	判别力	"Precision-Recall and ROC curves across model stages; AUC rises from 0.984 to 0.995."
Fig 6 Feature importance	§6.3	可解释	"Grouped feature importance; higher-order propagation and random-walk blocks dominate."
Fig 7 Error & robustness	§6	冷启动+决策	"(a) F1 by author/paper degree (cold-start drops); (b) calibration; (c) rank-cutoff vs probability-threshold stability."

附录图:USE embedding UMAP(度数着色)、14-LightGCN 相关性热图、超参热图、RW ensemble size vs F1。

---

6. Research Questions 与实验对应(advice.md 建议加 RQ)

RQ	问题	对应实验/证据
RQ1	不同边类型/传播阶对推荐性能的贡献?	高阶消融(Table 4):A-P-P^k vs A-A-P-P^k;fwd/bwd/undir
RQ2	冷启动作者能否从图结构获益?	误差分桶(Table 5):低度作者靠 has_local_evidence 与合著传播
RQ3	硬负采样 vs 随机负采样?	LightGCN 训练采用混合硬负采样(random/popular/coauthor 50/25/25)
RQ4	概率阈值 vs rank 决策的稳健性?	Table 6 + ratio-F1 曲线
RQ5	多源特征堆叠 vs 单 GNN?	Table 2 主结果

---

7. 写作注意事项(reviewer 视角)

1. 防 label leakage——必须明确写出验证边不进入 LightGCN 消息传递图,引用/合著网络是静态的(advice.md 重点)。
2. 评估协议——说明验证集是人工 1:1;F1 对阈值敏感,故补充 Precision/Recall/AUC(advice.md 建议)。
3. **不要只写"用了更高级模型"**——强调"根据任务特点选模型"(LightGCN 比 HGT 在小数据上更稳,呼应 advice.md)。
4. meta-path 明确化——把 A-A-P、A-P-P 写成正式 meta-path 术语,是异构图推荐的关键词。
5. 诚实标注局限——数据匿名无领域标签;top-50% 是基于榜榜反馈的假设;未来可加校准(Platt/Isotonic)与图注意力。
6. 复现性——所有结果 CPU 可复现,代码与缓存随包提供;F1 跨种子方差 <0.001。
7. 数字一致性——论文里所有 F1 必须与 validation_summary.csv / dynamic_summary.csv 一致;最终 0.9669(val)/0.9663(LB)。

---

8. 中英术语对照

中文	英文(论文用)
作者-论文链接预测	author–paper link prediction
异构图	heterogeneous graph
合著网络	co-authorship network
引用网络	citation network
二部图	bipartite graph
冷启动	cold start
矩阵分解	matrix factorization (MF)
随机游走	random walk (DeepWalk / Node2Vec)
高阶传播	higher-order propagation
元路径	meta-path
堆叠/二级学习器	stacking / second-stage (meta) learner
负采样	negative sampling (hard negatives)
秩截断决策	rank-cutoff decision rule
校准	calibration
分布漂移	distribution shift
消融	ablation (study)
特征重要性	feature importance

---

附录 A:关键文件速查(写论文/核对数据用)

用途	文件
最终成绩	validation_runs/dynamic_seed202/high_order_graph_stack/validation_summary.csv
LightGCN 超参搜索	validation_runs/dynamic_summary.csv
堆叠消融	validation_runs/dynamic_seed202/post95_ablation/ablation_table.csv
内容/BPR 消融	validation_runs/dynamic_seed202/extra_score_sources/extra_score_ablation.csv
随机游走消融	validation_runs/dynamic_seed202/randomwalk_systematic/{small,graph}_ablation_table.csv、ensemble_{5,7}_ablation.csv
误差分桶	validation_runs/dynamic_seed202/error_group_calibration/error_analysis_buckets.csv
ratio-F1 稳健性	validation_runs/stack_ratio_analysis.csv
决策规则漂移	validation_runs/dynamic_seed202/high_order_graph_stack/threshold_submission_summary.csv
实验演进叙事	notes/experiment_history.md、reports/{preliminary_report,exploration_summary,final_report}.md
最终提交	validation_runs/dynamic_seed202/high_order_graph_stack/submissions/submission_rich_rw7_highorder_directed_r0.500000.csv