Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Enhancing Contrastive Link Prediction With Edge Balancing Augmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that link prediction with contrastive learning is governed by the minimum node degree inside each latent cluster, and that raising that minimum via edge augmentation improves prediction.

desk verdict The central theory contradicts itself: Lemma 2's bound loosens when d̂_min increases, so Corollary 1 is backwards, yet the EBA method itself is a concrete, possibly useful empirical idea. read the letter →

arxiv 2508.14808 v1 pith:K7G6X4SR submitted 2025-08-20 cs.LG

Chen-Hao Chang , Hui-Ju Hung , Chia-Hsun Lu , Chih-Ya Shen This is my paper
classification cs.LG
keywords linkpredictiongraphcontrastivelearningaugmentationnodedegreeautoencoderembeddingconcentrationtheoreticalanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish the first theoretical account of how contrastive learning helps link prediction, and to extract a design rule from it. The rule is: the error rate of a contrastive autoencoder-based link predictor is bounded by a concentration term that shrinks as the minimum node degree within each latent cluster grows. On that basis the authors propose Edge Balancing Augmentation (EBA), which prunes low-confidence edges and adds edges to the most similar nodes, and CoEBA, which combines EBA with contrastive losses that treat one-hop neighbours as positive samples. If the account is right, degree-aware augmentation is not a heuristic; it follows from the bound, and it can be bolted onto existing graph autoencoders to improve Hits@10. Experiments on eight datasets report consistent gains, with a measured correlation between minimum degree and Hits@10.

What carries the argument

The load-bearing object is the (α,γ,d̂)-augmentation concentration bound: for each latent cluster there is a core subset whose diameter is at most γ(B/d̂_min)^{1/2}. Corollary 1 turns this into a design target—raise d̂_min to shrink the bound. The machinery that carries the argument is EBA: a warm-up encoder scores edges, low-confidence ones are pruned at ratio r_m, and each node is linked to the r_a fraction of most similar nodes, which raises the minimum degree while preserving structure; a random feature mask augments attributes. Around this sit the neighbor-concentrated contrastive losses, which use one-hop neighbours as positives in the original view and only the node itself as positive

What would settle it

Take a graph whose edges mostly cross communities (a heterophily benchmark) and run CoEBA against the same model with random edge additions matched to EBA's degree increase. If Hits@10 does not rise with the minimum degree—or rises no more than under random addition—then the bound's degree term is not the controlling mechanism. Alternatively, measure per-cluster minimum degree and per-cluster link-prediction error; Corollary 1 predicts a negative correlation, and its absence would falsify the claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that contrastive link prediction on autoencoder-based models inherits a concentration bound from node-classification theory, and that the actionable quantity in that bound is the minimum node degree d̂_min inside each latent cluster. The authors define the Potential-Link Concentration (PLCon) property—nodes that are linked tend to cluster—and argue from the reconstruction and contrastive losses that it holds. They then bound the similarity between embeddings from different clusters; the bound contains γ(B/d̂_min)^{1/2}. Corollary 1 states that improving link prediction means minimizing this term, which means increasing d̂_min (or reducing feature dimension B). EB

Load-bearing premise

The assumption that carries the argument is that concentration bounds proven for node classification transfer unchanged to link prediction, together with the unproven Potential-Link Concentration property that linked nodes cluster; if either fails, the bound that motivates EBA does not apply.

Editorial extensions

If this is right

  • Higher minimum node degree within each latent cluster yields more concentrated embeddings and a lower upper bound on link prediction error.
  • Graph augmentations for contrastive link prediction should be judged by how they change the minimum degree, not just by how much structure they disturb.
  • EBA can be applied as a plug-in to common autoencoder link predictors (the paper demonstrates GAE, GNAE, and VGNAE) and improves their Hits@10.
  • Contrastive losses for link prediction should treat neighbours as positive pairs; the ablation study reports that removing this treatment hurts performance.
  • The measured Pearson correlation between minimum node degree and Hits@10 (0.83 on Cora, 0.69 on Citeseer, 0.68 average) gives a direct empirical handle on the theory.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Going beyond the paper: if Corollary 1 is the true mechanism, then any augmentation that raises d̂_min without adding cross-cluster edges—for instance, adding edges among nodes already close in the latent space by a degree-matched random rule—should also improve performance; that is testable against EBA's similarity-based selection.
  • The same bound suggests shrinking feature dimension B would help as much as raising degrees, so feature-space dimensionality reduction is a second, unexploited lever.
  • The PLCon assumption limits the regime: in graphs where edges mostly run between communities (heterophily or bipartite structures), concentrating linked nodes could hurt, so the method's boundary is a graph where links imply similarity.
  • The theory is stated for autoencoders, but the degree-concentration argument is agnostic to the encoder; a natural extension is to test EBA-style degree balancing in subgraph-based link predictors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes CoEBA, a contrastive link prediction framework built on a graph autoencoder, with Edge Balancing Augmentation (EBA) that adds/removes edges based on predicted link confidence and embedding similarity, plus neighbor-concentrated contrastive losses. The paper claims a first formal theoretical analysis for contrastive learning on link prediction (Sec. 3.3), centered on Lemma 2 and Corollary 1, which states that increasing the minimum node degree d̂_min tightens the embedding concentration bound γ√(B/d̂_min) and improves link prediction. Experiments compare CoEBA to 10 baselines on 8 datasets.

Significance. If the theoretical analysis were sound, the paper would make a meaningful contribution: a degree-based augmentation principle derived from concentration bounds, with a plug-and-play module for autoencoder link predictors. The empirical component is substantial (8 datasets, ablations, integration with GAE/GNAE/VGNAE, efficiency results). However, the central theoretical claim is internally inconsistent with the paper's own Lemma 2, and the 'formal' analysis relies on unproven assumptions transferred from node-classification bounds. Since EBA is explicitly motivated by Corollary 1, the theory cannot be credited as the basis for the method.

major comments (3)
  1. [Sec. 3.3, Corollary 1 vs Lemma 2 (Eq. 5)] The sign of Corollary 1 is reversed relative to Lemma 2. Let x = γ√(B/d̂_min). In Lemma 2, ρ_max contains +Mαx/r and E_ij contains −Mx/(4r) inside terms that are squared or linear. Both the first term r²(1−ρ_max−√(2ρ_max−Δμ/2)) and E_ij are strictly decreasing in x. Since x decreases as d̂_min increases, increasing d̂_min makes the inter-cluster similarity upper bound larger, not smaller. Thus Corollary 1, which instructs increasing d̂_min to reduce the bound, contradicts the only formal result connecting d̂_min to link prediction. This is an internal inconsistency, not merely a missing proof, and it invalidates the theoretical motivation for EBA.
  2. [Sec. 3.3, Lemma 1] Lemma 1 is not proven; the argument assumes the conclusion. The proof states 'According to Def. 1, we expect an ideal encoder...' and then postulates a threshold η, without showing that training with L_recon and L_contrastive yields a finite δ for connected nodes and a larger distance for unconnected nodes. The PLCon property is therefore assumed, not demonstrated. Since Lemma 2 and Corollary 1 both rely on this clustering behavior, the theoretical chain is unsupported at its base.
  3. [Sec. 3.3, Eq. (3) and Lemma 2] The bounds imported from [34] and [12] are derived for node classification under (α,γ,d̂)-augmentation. The paper transfers them to link prediction without verifying that the augmentation distribution over potential links satisfies the same conditions, and it does not define the analogue of 'cluster' for link prediction beyond the assumed PLCon property. Additionally, the final expression in Eq. (5) is assembled by combining a cluster-center bound with a deviation bound that is squared; the squared term requires an assumption about equal or worst-case deviation norms that is not stated. Corollary 1 is asserted without proof. The claim of 'first formal theoretical analysis' is therefore not supported.
minor comments (5)
  1. [Sec. 3.3, Eq. (5)] Equation (5) contains malformed parentheses and OCR-like artifacts (e.g., '!!'), and the definition of Δμ is ambiguous. Please rewrite the expression cleanly and define all symbols in one place.
  2. [Sec. 4.1] The range for r_a is written as '[1, 0]', which is presumably a typo for '[0, 1]'. Also, Corollary 1 mentions decreasing B (feature dimension) as a design option, but B is fixed for a given dataset; this is not an actionable augmentation knob.
  3. [Sec. 5, Pearson correlation paragraph] The reported Pearson correlations (0.83 on Cora, 0.69 on Citeseer, average 0.68) are not documented: over what set of models or augmentation strengths are these computed? No error bars or scatter data are provided. As stated, this correlation is not evidence for the causal claim that increasing d̂_min improves link prediction.
  4. [Sec. 5, Figure 3] The sensitivity plot shows Hits@10 values but no standard deviations or axis labels for the color bar; please add details and clarify whether the reported values are averaged over multiple splits.
  5. [General] The reproducibility statement says materials are 'available upon request'; for a CIKM paper, a public code repository would be expected.

Circularity Check

2 steps flagged · score 4.0 of 10

The theory is partially circular: Lemma 1 restates the loss objectives as a discovered clustering property, and Corollary 1 reads d̂_min off Definition 1 as a design knob; the empirical evaluation remains independent.

  1. self definitional [Sec. 3.3, Lemma 1 and its proof]
    "With the reconstruction loss in Eq. 1, the model maximizes E_q(Z|X,A)[log p(A|Z)] ... Since we predict A_pred by calculating the inner product of Z, σ(z_i^T z_j), therefore, the model maximizes z_i^T z_j when A_ij = 1 and minimizes z_i^T z_j when A_ij = 0. As for the contrastive loss in Eq. 2 ... Maximizing the similarity of a pair of positive samples v_i and v_j increases z_i^T z_j."

    Lemma 1 is the load-bearing step that establishes the PLCon property (nodes with potential links cluster). But its proof only describes the training objectives: Eq. 1 is the reconstruction loss that rewards high inner products for edges, and Eq. 2 is the contrastive loss that rewards high similarity for positive pairs. The conclusion that connected nodes are closer in latent space is therefore already contained in the construction of the losses. The paper then treats this restatement of the loss design as a formal guarantee that contrastive learning induces the clustering needed for link prediction, which is circular: the property is put in by the loss design, not derived for the link-prediction setting.

  2. self definitional [Sec. 3.3, Corollary 1]
    "Corollary 1. To improve the performance of the link prediction model employing the contrastive loss, enhancing the concentration of embeddings is critical. This involves minimizing the upper bound γ(B/d̂_min)^{1/2} in sup_{v_i,v_j∈C^0_p} d_T(v_i,v_j) ≤ γ(B/d̂_min)^{1/2}, where γ∈(0,1]. Essentially, this requires decreasing B or increasing d̂_min."

    The corollary is not obtained from Lemma 2; as written, Lemma 2's right-hand side increases when d̂_min increases, moving opposite to Corollary 1. Instead, Corollary 1 restates Definition 1, where the (α,γ,d̂)-augmentation is defined with exactly the bound d_T ≤ γ√(B/d̂_min) and d̂_min in the denominator. Since Definition 1 is an imported definition (from [34]), the instruction to 'increase d̂_min' is a direct read-off of that definition, not an independent link-prediction theorem. EBA is then constructed to implement this read-off, so the theoretical motivation and the augmentation design are the same quantity.

full rationale

The paper has an extensive and self-contained empirical component: CoEBA is compared against 10 baselines on 8 datasets, with ablations, plug-and-play integration with GAE/GNAE/VGNAE, and efficiency studies. These results do not reduce to the theoretical claims, so the paper is not wholly circular. The self-citations present (e.g., [39]) are not load-bearing for the theory. The main circularity is confined to the theoretical motivation: Lemma 1 is a restatement of the reconstruction/contrastive objectives, and Corollary 1 is a restatement of Definition 1's concentration bound with d̂_min as the design knob. Also, the paper's own Lemma 2 moves in the opposite direction from Corollary 1 as written, so the corollary is not a derived consequence of the preceding equations; it is an assertion that re-imports the definition. This makes the 'first formal theoretical analysis' claim substantially weaker, but because the empirical evaluation is independent and no fitted parameter is renamed as a prediction, the appropriate circularity score is moderate, not extreme.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unproved imports: the entire (α,γ,d̂)-augmentation framework of Wang et al. [34], the deviation bound of Huang et al. [12], and the assumption that these bounds apply to link prediction. The only parameters fitted to the benchmark data are the hyperparameters (r_m, r_a, λs), which directly control the augmentation whose benefit the theory is supposed to predict. No new physical or mathematical entities are introduced.

free parameters (6)
  • neighbor removal ratio r_m = 0.14
    Chosen by grid search on Cora (Fig. 3) and used across datasets; directly controls EBA edge pruning.
  • neighbor addition ratio r_a = 0.40
    Chosen by grid search on Cora (Fig. 3); controls the number of new edges added per node.
  • loss weights λ1, λ2, λ3 = 3, 1, 3
    Grid search in Eq. 12; balance reconstruction and contrastive losses.
  • feature masking ratio = not specified
    Random feature mask is applied following [50], but the ratio is not reported.
  • temperature τ in contrastive losses = not specified
    Used in Eqs. 9-11 but its value is not stated.
  • augmentation update interval t = not specified
    Algorithm 1 uses 'if e % t == 0' but t is never defined in the text.
assumptions (4)
  • domain assumption The (α,γ,d̂)-augmentation definition and Theorem 1 from Wang et al. [34] hold and transfer to link prediction.
    Invoked in Sec. 3.2 and used to build Lemma 2; applicability to link prediction is never proven.
  • domain assumption The deviation bound from Huang et al. [12] (Eq. 3) holds for the link prediction setting without modification.
    Used directly in the proof of Lemma 2 to bound ||e_i||; the bound was derived for node classification.
  • domain assumption Embeddings are normalized with a fixed norm r (or 1), and the distance threshold δ is finite [26].
    Used in the proofs of Lemma 1 and Lemma 2; the finite δ is cited from prior work.
  • ad hoc to paper The Potential-Link Concentration (PLCon) property holds: nodes sharing potential links are more likely to be in the same latent cluster.
    Introduced in Sec. 3.3 as the basis for the whole analysis, but not proven; it is assumed that the losses induce this clustering for missing edges.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Contrastive Link Prediction With Edge Balancing Augmentation." pith.science (2026). https://pith.science/paper/K7G6X4SR

@misc{pith2026250814808,
  author       = {Pith},
  title        = {Pith review of: Enhancing Contrastive Link Prediction With Edge Balancing Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7G6X4SR}},
  note         = {Machine review of arXiv:2508.14808}
}
read the original abstract

Link prediction is one of the most fundamental tasks in graph mining, which motivates the recent studies of leveraging contrastive learning to enhance the performance. However, we observe two major weaknesses of these studies: i) the lack of theoretical analysis for contrastive learning on link prediction, and ii) inadequate consideration of node degrees in contrastive learning. To address the above weaknesses, we provide the first formal theoretical analysis for contrastive learning on link prediction, where our analysis results can generalize to the autoencoder-based link prediction models with contrastive learning. Motivated by our analysis results, we propose a new graph augmentation approach, Edge Balancing Augmentation (EBA), which adjusts the node degrees in the graph as the augmentation. We then propose a new approach, named Contrastive Link Prediction with Edge Balancing Augmentation (CoEBA), that integrates the proposed EBA and the proposed new contrastive losses to improve the model performance. We conduct experiments on 8 benchmark datasets. The results demonstrate that our proposed CoEBA significantly outperforms the other state-of-the-art link prediction models.

Figures

Figures reproduced from arXiv: 2508.14808 by the authors.

Figure 1
Figure 1. Illustrations of Def. 1 and Def. 2. (a) Augmentation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed CoEBA: (a) Overview of the full pipeline; (b) Illustration of Edge Balancing Augmentation (EBA); (c) The proposed neighbor-concentrated contrastive objective obtain the initial model and the corresponding node embeddings. This ensures that the node embeddings are adequately refined and stable, making them suitable to perform EBA. We first obtain the embedding 𝑍 by inputting 𝑋 and 𝐴 to th… view at source ↗
Figure 3
Figure 3. Sensitivity tests of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 43 canonical work pages

  1. [34]

    Ruijia Wang, Xiao Wang, Chuan Shi, and Le Song. 2022. Uncovering the structural fairness in graph contrastive learning. InNeurIPS. 32465–32473

    work page 2022

  2. [12]

    Weiran Huang, Mingyang Yi, Xuyang Zhao, and Zihao Jiang. 2023. Towards the Generalization of Contrastive Self-Supervised Learning. InICLR

    work page 2023

  3. [1]

    Seong Jin Ahn and Myoung Ho Kim. 2021. Variational Graph Normalized Autoen- coders. InCIKM

    work page 2021

  4. [2]

    Carnegie Mellon University. 2001. WebKB Project. https://www.cs.cmu.edu/project/ theo-11/www/wwkb/

    work page 2001

  5. [3]

    Bronstein, and Max Hansmire

    Benjamin Paul Chamberlain, Sergey Shirobokov, Emanuele Rossi, Fabrizio Frasca, Thomas Markovich, Nils Hammerla, Michael M. Bronstein, and Max Hansmire

    work page

  6. [4]

    Jialu Chen and Gang Kou. 2023. Attribute and structure preserving graph con- trastive learning. InAAAI. 7024–7032

    work page 2023

  7. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A Simple Framework for Contrastive Learning of Visual Representations. InICML

    work page 2020

  8. [6]

    Wenqi Fan, Yao Ma, Qing Li, Jianping Wang, Guoyong Cai, Jiliang Tang, and Dawei Yin. 2022. A Graph Neural Network Framework for Social Recommendations.IEEE Transactions on Knowledge and Data Engineering34, 5 (2022), 2033–2047

    work page 2022

Show all 53 references
  1. [7]

    Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Gnnemann. 2019. Com- bining Neural Networks with Personalized PageRank for Classification on Graphs. InICLR

    2019

  2. [8]

    C Lee Giles, Kurt D Bollacker, and Steve Lawrence. 1998. CiteSeer: An automatic citation indexing system. InProceedings of the third ACM conference on Digital libraries. 89–98

    1998

  3. [9]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. 2017. Neural Message Passing for Quantum Chemistry. InICML. 1263–1272

    2017

  4. [10]

    Zhihao Guo, Feng Wang, Kaixuan Yao, Jiye Liang, and Zhiqiang Wang. 2022. Multi- Scale Variational Graph AutoEncoder for Link Prediction. InWSDM

    2022

  5. [11]

    Selim Haj Ali and Marc-Thorsten Hütt. 2022. Inferring missing edges in a graph from observed collective patterns.Physical Review E105, 6 (2022), 064610

    2022

  6. [13]

    Zexi Huang, Mert Kosan, Arlei Silva, and Ambuj Singh. 2023. Link prediction without graph neural networks.arXiv preprint arXiv:2305.13656(2023)

    2023 arXiv

  7. [14]

    Junghun Kim, Ka Hyun Park, Hoyoung Yoon, and U Kang. 2025. Accurate Link Prediction for Edge-Incomplete Graphs via PU Learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 17877–17885

    2025

  8. [15]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Optimiza- tion.arXiv preprint arXiv:1412.6980(2017)

    2017 arXiv

  9. [16]

    Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders.arXiv preprint arXiv:1611.07308(2016)

    2016 arXiv

  10. [17]

    Gueorgi Kossinets. 2006. Effects of missing data in social networks.Social networks 28, 3 (2006), 247–268

    2006

  11. [18]

    Sanjoy Kundu and Sathyanarayanan N. Aakur. 2023. IS-GGT: Iterative Scene Graph Generation With Generative Transformers. InCVPR. 6292–6301

    2023

  12. [19]

    Juanhui Li, Harry Shomer, Haitao Mao, Shenglai Zeng, Yao Ma, Neil Shah, Jiliang Tang, and Dawei Yin. 2024. Evaluating graph neural networks for link prediction: current pitfalls and new benchmarking. InNeurIPS

    2024

  13. [20]

    Wen-Zhi Li, Chang-Dong Wang, Jian-Huang Lai, and Philip S. Yu. 2024. Towards Effective and Robust Graph Contrastive Learning With Graph Autoencoding.IEEE Transactions on Knowledge and Data Engineering36, 2 (2024), 868–881. https: //doi.org/10.1109/TKDE.2023.3288280

    2024

  14. [21]

    Ke Liang, Yue Liu, Sihang Zhou, Wenxuan Tu, Yi Wen, Xihong Yang, Xiangjun Dong, and Xinwang Liu. 2023. Knowledge graph contrastive learning based on relation- symmetrical structure.IEEE Transactions on Knowledge and Data Engineering36, 1 (2023), 226–238

    2023

  15. [22]

    Li Ma, Haoyu Han, Juanhui Li, Harry Shomer, Hui Liu, Xiaofeng Gao, and Jiliang Tang. 2024. Mixture of Link Predictors on Graphs. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  16. [23]

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore

  17. [24]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    2018 arXiv

  18. [25]

    Jongwon Park, Heesoo Jung, and Hogun Park. 2025. CIMAGE: Exploiting the Conditional Independence in Masked Graph Auto-encoders. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 10–19

    2025

  19. [26]

    Purnamrita Sarkar, Deepayan Chakrabarti, and Andrew W Moore. 2011. Theoretical Justification of Popular Link Prediction Heuristics.. InIJCAI

    2011

  20. [27]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of graph neural network evaluation.arXiv preprint arXiv:1811.05868(2018)

    2018 arXiv

  21. [28]

    Xiao Shen, Dewang Sun, Shirui Pan, Xi Zhou, and Laurence T Yang. 2023. Neighbor contrastive learning on learnable graph augmentation. InAAAI. 9782–9791

    2023

  22. [29]

    Papalexakis, Yozen Liu, and Neil Shah

    William Shiao, Zhichun Guo, Tong Zhao, Evangelos E. Papalexakis, Yozen Liu, and Neil Shah. 2023. Link Prediction with Non-Contrastive Learning. InICLR

    2023

  23. [30]

    Qiaoyu Tan, Ninghao Liu, Xiao Huang, Soo-Hyun Choi, Li Li, Rui Chen, and Xia Hu. 2022. S2GAE: Self-Supervised Graph Autoencoders are Generalizable Learners with Graph Masking. InWSDM

    2022

  24. [31]

    Zhiquan Tan, Yifan Zhang, Jingqin Yang, and Yang Yuan. 2024. Contrastive Learning is Spectral Clustering on Similarity Graph. InICLR

    2024

  25. [32]

    Yuandong Tian. 2022. Understanding Deep Contrastive Learning via Coordinate- wise Optimization. InNeurIPS, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.). 19511–19522

    2022

  26. [33]

    Yuandong Tian. 2023. Understanding the Role of Nonlinearity in Training Dynamics of Contrastive Learning. InICLR

    2023

  27. [35]

    Tongzhou Wang and Phillip Isola. 2020. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. InICML. 9929– 9939

    2020

  28. [36]

    Xiyuan Wang, Haotong Yang, and Muhan Zhang. 2024. Neural Common Neighbor with Completion for Link Prediction. InICLR

    2024

  29. [37]

    Yu Wang, Yuying Zhao, Yi Zhang, and Tyler Derr. 2023. Collaboration-Aware Graph Convolutional Network for Recommender Systems. Inthe Web Conference. 91–101

    2023

  30. [38]

    Teng Xiao, Huaisheng Zhu, Zhengyu Chen, and Suhang Wang. 2023. Simple and Asymmetric Graph Contrastive Learning without Augmentations. InThirty-seventh Conference on Neural Information Processing Systems

    2023

  31. [39]

    Hao-Wei Yang, Ming-Yi Chang, and Chih-Ya Shen. 2023. Enhancing Link Prediction with Self-Discriminating Augmentation for Structure-Aware Contrastive Learning. InECAI. 2842–2849

    2023

  32. [40]

    Lichtenwalter, and Nitesh V

    Yang Yang, Ryan N. Lichtenwalter, and Nitesh V. Chawla. 2015. Evaluating link prediction methods.Knowledge and Information Systems45, 3 (2015), 751–782

    2015

  33. [41]

    Yihang Yin, Qingzhong Wang, Siyu Huang, Haoyi Xiong, and Xiang Zhang. 2022. AutoGCL: Automated Graph Contrastive Learning via Learnable View Generators. InAAAI

    2022

  34. [42]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph Contrastive Learning with Augmentations. InNeurIPS

    2020

  35. [43]

    Seongjun Yun, Seoyoon Kim, Junhyun Lee, Jaewoo Kang, and Hyunwoo J. Kim

  36. [44]

    Lorenzo Zangari, Domenico Mandaglio, and Andrea Tagarelli. 2024. Link prediction on multilayer networks through learning of within-layer and across-layer node- pair structural features and node embedding similarity. InProceedings of the ACM Web Conference 2024. 924–935

    2024

  37. [45]

    Muhan Zhang and Yixin Chen. 2018. Link Prediction Based on Graph Neural Networks. InNeurIPS

    2018

  38. [46]

    Zehua Zhang, Shilin Sun, Guixiang Ma, and Caiming Zhong. 2023. Line graph contrastive learning for link prediction.Pattern Recognition140 (2023), 109537

    2023

  39. [47]

    Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, and Meng Jiang. 2022. Learning from Counterfactual Links for Link Prediction. InICML. 26911–26926

    2022

  40. [48]

    Zhanke Zhou, Jiangchao Yao, Jiaxu Liu, Xiawei Guo, quanming yao, LI He, Liang Wang, Bo Zheng, and Bo Han. 2023. Combating Bilateral Edge Noise for Robust Link Prediction. InThirty-seventh Conference on Neural Information Processing Systems. CIKM ’25, November 10–14, 2025, Seou...

    2023

  41. [49]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep graph contrastive representation learning.arXiv preprint arXiv:2006.04131(2020)

    2020 arXiv

  42. [50]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph Contrastive Learning with Adaptive Augmentation. Inthe Web Conference

    2021

  43. [2000]

    Information Retrieval3 (2000), 127–163

    Automating the construction of internet portals with machine learning. Information Retrieval3 (2000), 127–163

    2000

  44. [2022]

    InNeurIPS

    Neo-GNNs: Neighborhood Overlap-aware Graph Neural Networks for Link Prediction. InNeurIPS

  45. [2023]

    Graph Neural Networks for Link Prediction with Subgraph Sketching. In ICLR

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.