Pith. sign in

REVIEW 4 major objections 6 minor 61 references

Str-GCL: Structural Commonsense Driven Graph Contrastive Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes Str-GCL, a graph contrastive learning framework that encodes two pieces of structural commonsense—about neighbor degrees and feature distinctiveness—as first-order logic rules, and shows that aligning encoder…

desk verdict Worth reviewing, but the structural-commonsense rules were selected on the same benchmarks used for evaluation; out-of-sample validation is needed before the generalizability claim holds. read the letter →

arxiv 2507.07141 v1 pith:HU3ZLY74 submitted 2025-07-09 cs.LG

Dongxiao He , Yongqi Huang , Jitao Zhao , Xiaobao Wang , Zhen Wang This is my paper
classification cs.LG
keywords GraphContrastiveLearningStructuralCommonsenseFirst-OrderLogicRulesNodeClassificationSelf-SupervisedRepresentationAlignmentNeuralNetworks
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

Current graph contrastive learning methods, the paper argues, learn only implicit semantic relationships and therefore keep misclassifying a stable set of nodes. The paper's claim is that part of what those models miss is structural commonsense—simple, human-intuitive facts about local topology and feature distinctiveness—and that this commonsense can be written down as first-order logic rules and fed into training. Str-GCL implements this by building a second, rule-driven representation from the original graph without extra augmentations and aligning the encoder's node representations to it in mean and covariance, together with a dedicated rule-space contrastive loss. The reported result is that Str-GCL exceeds existing GCL baselines on six benchmark graphs in node classification, improves clustering, and works as a drop-in plugin that raises the accuracy of GRACE, CCA-SSG, and DGI. A sympathetic reader should care because it offers a path to self-supervised graph learning that is both more accurate and more interpretable without any extra labels.

What carries the argument

A pair of rule scorers derived from first-order-logic-style constraints: NTSC (Neighborhood Topological Summation Constraint) upweights nodes whose neighbors have a small total degree, on the premise that averaging over few low-degree neighbors does not smooth training noise; LGTC (Local-Global Threshold Constraint) upweights nodes whose neighbor feature similarity barely exceeds their global feature similarity, on the premise that such nodes lack class-distinctive features. These rule weights gate a rule representation built by an MLP over PCA-reduced original node features, and the mechanism that bridges rules to the encoder is a distributional alignment loss: it matches the mean vector and covariance matrix of the rule representation to those of the node representations, plus a contrastive loss that shapes the rule space itself. The ablations show that removing the alignment loss costs more than removing the rule-space loss, so the alignment step is the main carrier of the structural commonsense.

What would settle it

Run Str-GCL on graphs that were not used in rule discovery—for instance heterophilic benchmarks such as Chameleon or Squirrel, or the six datasets with randomly permuted node features—and check two things: whether NTSC/LGTC-flagged nodes are disproportionately misclassified, and whether Str-GCL still beats GRACE. If flagged nodes are not harder than average, or the gains vanish, the claim that the rules express reusable structural commonsense is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a large share of consistently misclassified nodes in graph contrastive learning are not arbitrary failures but nodes that violate two simple structural patterns: nodes whose first-order neighbors have a low total degree are poorly averaged and therefore unstable, and nodes whose average feature similarity to their neighbors is close to their average similarity to the whole graph lack class-distinctive features. Str-GCL encodes these two intuitions as first-order logic rules—NTSC for the neighborhood degree summation and LGTC for the local-global similarity gap—converts them into per-node attention weights, and produces a rule representation from PCA-reduced original features processed by an MLP and gated by those weights. The encoder is then trained with the usual InfoNCE contrastive loss plus a contrastive loss in rule space and a cross-representation alignment loss that matches the mean and covariance of rule and node representations, which the ablations show is the component that transmits the commonsense to the node embeddings; on six datasets the reported outcome is better node classification than all compared GCL baselines, improved clustering on Cora and CS, and accuracy gains when the rule module is added to GRACE, CCA-SSG, and DGI.

Load-bearing premise

The two rules were inferred from misclassification statistics on the same six benchmark graphs used to measure success, so the central assumption is that they encode general structural commonsense rather than dataset-specific artifacts.

Editorial extensions

If this is right

  • Str-GCL posts higher node-classification accuracy than every compared GCL baseline on Cora, CiteSeer, PubMed, CS, Photo, and Computers, and also surpasses supervised GCN on all six.
  • Attached as a plugin, Str-GCL raises the accuracy of GRACE, CCA-SSG, and DGI on each dataset, with gains up to 2.0 points on Photo and Computers for GRACE.
  • Frequent misclassifications fall after adding the rules: the count of nodes wrong in 15 or more of 20 runs drops by 4.44% to 16.68% across datasets, matching the paper's motivation about error-prone nodes.
  • On clustering, Str-GCL records the best NMI and ARI on Cora and CS among the compared methods, with CS ARI jumping from 0.654 (GRACE) to 0.785.
  • The alignment component is essential: ablating Lcross consistently costs more accuracy than ablating Lrule alone.

Reading between the lines

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

  • Because NTSC and LGTC were discovered on the same benchmarks used for evaluation, the rules' generality is untested; deriving the rules on disjoint graphs and transferring them would separate true commonsense from dataset-specific error patterns.
  • The attention weights act as a hard-example mining signal, so a natural extension is to replace the hand-written rules with a learned error-proneness scorer and ask whether the alignment mechanism still transfers.
  • Both rules assume homophily: on heterophilic graphs where neighbors tend to belong to different classes, neighbor-degree averaging and local similarity are likely inverted signals, so a heterophily-aware rule variant is a plausible next step.
  • The covariance-plus-mean alignment is a simple distributional distillation from rule space to node space; comparing it with adversarially matched or KL-based alignment would show whether the specific MSE form is what carries the gain.
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

4 major / 6 minor

Summary. The paper proposes Str-GCL, a graph contrastive learning method that injects two human-interpretable rules, NTSC and LGTC, into the training objective. NTSC up-weights nodes with low aggregate neighbor degree; LGTC up-weights nodes whose local feature similarity is close to their global similarity. Rule representations are generated from PCA-projected features and aligned with node representations via mean and covariance matching losses. Experiments on Cora, CiteSeer, PubMed, Coauthor CS, Amazon Photo, and Amazon Computers report consistent improvements over GCL baselines in node classification and clustering, and show that Str-GCL can be added as a plugin to GRACE, CCA-SSG, and DGI.

Significance. If valid, Str-GCL would be a useful, interpretable mechanism for incorporating structural priors into contrastive graph learning, with an unusually thorough experimental section: six datasets, classification and clustering, ablations, plugin experiments, and error-distribution analysis. However, the core generalizability claim is currently undermined by the provenance of the rules, which were discovered through error analysis on the same datasets used for evaluation, and by invalid supporting proofs in Appendix A.1 and A.2. The experimental surface is solid, but the conceptual contribution needs additional validation before the claimed general structural commonsense can be accepted.

major comments (4)
  1. [§2, §4.1, Tables 1 and 3] The rule set is circularly selected on the evaluation data. Section 2 identifies consistently misclassified nodes by running GRACE 20 times on PubMed and CS and, in Appendix A.7, on Computers and Photo, and Section 4.1 states that NTSC and LGTC were obtained 'through observation and statistical analysis' of these error patterns. Because the same six datasets are then used to demonstrate the gains in Tables 1 and 3, the reported improvements are at least partly in-sample. The conclusion that Str-GCL captures generalizable structural commonsense would require validating the rules on graphs that played no role in their discovery, or pre-registering the rules before the error analysis.
  2. [Appendix A.1, Eqs. (15)-(18)] The proof that low TotalDegree causes instability is invalid. The total gradient in Eq. (15) accumulates the raw noise sum ∑ε_u; Eq. (17) divides that sum by TotalDegree, but the gradient is not normalized by neighbor-degree sum, so the comparison Noise_w/TotalDegree(w) < Noise_v/TotalDegree(v) does not imply that node w has smaller noise impact in its gradient. The law of large numbers is also invoked without the required independent, identically distributed and degree-invariant noise model. This leaves the NTSC rule without theoretical support.
  3. [Appendix A.2, Eqs. (20)-(23)] The LGTC proof relies on an invalid inference. From LocalSim(v) ≈ GlobalSim(v) the paper concludes x_u ≈ x̄ for every neighbor u in Eq. (22). Equality of average similarities does not imply that each neighbor feature equals the global mean; many feature configurations can yield equal averages. Without a valid derivation, the claim that a small local-global similarity gap implies indistinctive class features is unproven.
  4. [Table 1 and Table 3] Many of the reported gains are within one standard deviation of the strongest baseline (e.g., Cora 84.89±0.90 vs. PiGCL 84.63±0.78; PubMed 86.81±0.14 vs. PiGCL 86.75±0.20), and no significance tests or effect sizes are reported. The statement that Str-GCL 'significantly outperforms' existing methods should be supported by statistical testing or qualified accordingly.
minor comments (6)
  1. [§4.1, Eq. (3)] The formalization is not a well-formed first-order logic rule: the second line is an arithmetic definition rather than a formula, and the sentence 'Neighbor(v_i,v_j) represents the total sum' should refer to TotalDegree. Please rewrite the rule presentation to match the intended semantics.
  2. [§4.2, Eq. (9)] The notation Cov(A) and Cov(B) is undefined; these should be Cov(H_N) and Cov(H_R) as defined in Eq. (8).
  3. [Appendix A.2, Eq. (19)] GlobalSim(v) is defined as an average over u∈V, which includes v itself; the main text defines it over V\{v}. Please make the definitions consistent.
  4. [Throughout] There are several typos: Table 4 title 'Comparsion' should be 'Comparison'; Appendix A.1 contains 'sepcific' and 'an graph'; Section 6.7 header 'Visualizaion' should be 'Visualization'.
  5. [Algorithm 1] Line 2 says 'Calculate NTSC and LGTC, and get original features X_R' but the rules produce weights, not features; also lines 7-8 conflate rule representations H_R with learnable weights. Please clarify the pseudocode.
  6. [Appendix A.8] No source code or repository link is provided for Str-GCL itself; only links to baseline implementations are listed in Table 9. Making the code available would strengthen reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

NTSC and LGTC rule sets are selected from error-prone-node statistics on the same benchmarks used for evaluation, making the reported reduction in misclassified nodes partly in-sample.

  1. fitted input called prediction [Section 4.1 (rule derivation), with Section 2 (rule discovery) and Section 6.5/Table 4 (evaluation)]
    "Through observation and statistical analysis, we have identified that sets of nodes adhering to certain observable human patterns are more prone to misclassification than those outside these patterns. These statistically derived intuitions serve as a bridge between error-prone nodes and observed patterns."

    Section 2 defines error-prone nodes as those misclassified at least 15 times by GRACE across 20 runs on PubMed and CS, and Appendix A.7 extends this analysis to Photo and Computers. Section 4.1 then states that NTSC and LGTC were identified through observation and statistical analysis of such nodes, with weights w_i = max(hat-d_log) - hat-d_log(i) and s_i = max(Diff) - Diff(i) upweighting precisely the low-neighbor-degree and low-distinctiveness properties observed in that same node set. Section 6.5 and Table 4 report the reduction in nodes misclassified at least 15 times on the same PubMed, CS, Photo, and Computers benchmarks.

full rationale

The central circularity is not equation-level tautology but selection-on-evaluation-data: the two rules NTSC and LGTC are discovered by analyzing which nodes GRACE misclassifies repeatedly on PubMed, CS, Photo, and Computers, and the same four datasets are then used to claim that Str-GCL reduces exactly that class of frequently misclassified nodes. The paper does not evaluate the rules on any dataset that was not used in their derivation, so the claimed ability of the rules to capture generalizable structural commonsense is not established independently of the fit. This is a genuine fitted-input-called-prediction pattern, though it is partial rather than total: the method could still improve held-out nodes within those graphs, and the plugin results on CCA-SSG and DGI provide some, but not dataset-independent, evidence. The appendix proofs in A.1 and A.2 are mathematically shaky (e.g., deriving an inequality from an approximate equality), but that is a correctness concern, not a circularity concern. There is no load-bearing self-citation chain and no uniqueness argument imported from the authors' prior work; the paper's self-citations are to related GCL papers and are not the basis of the claimed prediction. Overall, the rule provenance makes a substantial part of the empirical claim in-sample, warranting a score of 6.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

No new physical or ontological entities are postulated; the rule representation branch is an architectural construct rather than a new entity. The central assumptions are the homophily prior, the GCN first-order neighborhood premise, and the two ad hoc statistical assumptions used in the appendix proofs. The many per-dataset hyperparameters are free choices that influence the reported gains.

free parameters (9)
  • InfoNCE temperature tau = 0.4 to 0.5 (Table 6)
    Chosen per dataset; controls contrastive sharpness and affects all reported results.
  • Rule loss temperature tau_rule = 0.3 to 0.7 (Table 6)
    Chosen per dataset; controls the rule representation contrastive loss.
  • Weight of Lrule = 100 on Cora, 1 on other datasets (Table 7)
    Hand-set per dataset; directly weights the rule loss term in the total objective.
  • Weight of Lcross = 1.0 (Table 7)
    Hand-set per dataset; weights the alignment loss between rule and node representations.
  • Learning rate = 1e-4 to 1e-2 (Table 6)
    Tuned per dataset for the reported performance.
  • Weight decay = 1e-5 to 5e-4 (Table 6)
    Tuned per dataset for the reported performance.
  • Number of training epochs = 800 to 18,000 (Table 6)
    Tuned per dataset; ranges widely across datasets.
  • Hidden dimension and MLP hidden dimension = 512 to 2048 and 32 to 512 (Table 6)
    Architecture sizes chosen per dataset.
  • Drop edge and drop feature rates = 0.1 to 0.4 (Table 7)
    Augmentation strengths tuned per dataset.
assumptions (6)
  • standard math Law of large numbers justifies noise averaging over a node's neighbor set
    Invoked in Appendix A.1 Eq. 17 to argue high total neighbor degree reduces gradient noise; the application is flawed because noise terms are not scaled by neighbor degrees.
  • domain assumption Homophily: neighbors tend to share class and features
    Section 4 focuses on homophilic graphs; both LGTC and the claim that feature similarity to neighbors signals class rely on this.
  • domain assumption First-order neighbors dominate GCN representations
    NTSC states first-order neighbors have the greatest impact; this is true of the GCN layer used but shapes both rules.
  • ad hoc to paper i.i.d. additive noise in gradient terms
    Eq. 14 postulates each neighbor gradient contribution equals true gradient plus independent noise; no support is given for the independence or identical distribution.
  • ad hoc to paper LocalSim near GlobalSim implies each neighbor feature equals the global mean
    Appendix A.2 Eq. 22; an average similarity being close to the global average does not imply individual neighbor vectors equal the mean.
  • ad hoc to paper Labels may be used in rule discovery though training is unsupervised
    Section 2 uses labels to count misclassifications of GRACE and Section 4.1 uses this analysis to choose rules; the paper frames the setting as unsupervised, which is only true of the training loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Str-GCL: Structural Commonsense Driven Graph Contrastive Learning." pith.science (2026). https://pith.science/paper/HU3ZLY74

@misc{pith2026250707141,
  author       = {Pith},
  title        = {Pith review of: Str-GCL: Structural Commonsense Driven Graph Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HU3ZLY74}},
  note         = {Machine review of arXiv:2507.07141}
}
read the original abstract

Graph Contrastive Learning (GCL) is a widely adopted approach in self-supervised graph representation learning, applying contrastive objectives to produce effective representations. However, current GCL methods primarily focus on capturing implicit semantic relationships, often overlooking the structural commonsense embedded within the graph's structure and attributes, which contains underlying knowledge crucial for effective representation learning. Due to the lack of explicit information and clear guidance in general graph, identifying and integrating such structural commonsense in GCL poses a significant challenge. To address this gap, we propose a novel framework called Structural Commonsense Unveiling in Graph Contrastive Learning (Str-GCL). Str-GCL leverages first-order logic rules to represent structural commonsense and explicitly integrates them into the GCL framework. It introduces topological and attribute-based rules without altering the original graph and employs a representation alignment mechanism to guide the encoder in effectively capturing this commonsense. To the best of our knowledge, this is the first attempt to directly incorporate structural commonsense into GCL. Extensive experiments demonstrate that Str-GCL outperforms existing GCL methods, providing a new perspective on leveraging structural commonsense in graph representation learning.

Figures

Figures reproduced from arXiv: 2507.07141 by the authors.

Figure 1
Figure 1. Two basic structural commonsense examples in gen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Misclassified nodes distribution of PubMed and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overview of the proposed method. Two graph views [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: T-SNE embeddings of nodes in Amazon Computers [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Misclassified nodes distribution comparison of [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [1]

    Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip S. Yu. 2023. Graph Self-Supervised Learning: A Survey. IEEE Trans. Knowl. Data Eng. 35, 6 (2023), 5879–5900

    work page 2023

  2. [2]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI Open 1 (2020), 57–81

    work page 2020

  3. [3]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net

    work page 2017

  4. [4]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. International Confer- ence on Learning Representations (2018)

    work page 2018

  5. [5]

    Yixin Liu, Kaize Ding, Qinghua Lu, Fuyi Li, Leo Yu Zhang, and Shirui Pan. 2023. Towards Self-Interpretable Graph-Level Anomaly Detection. In Advances in Neural Information Processing Systems 36: NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023

    work page 2023

  6. [6]

    Zhen Wang, Chunjiang Mu, Shuyue Hu, Chen Chu, and Xuelong Li. 2022. Mod- elling the Dynamics of Regret Minimization in Large Agent Populations: a Master Equation Approach.. In IJCAI. 534–540

    work page 2022

  7. [7]

    Yiqi Dong, Dongxiao He, Xiaobao Wang, Yawen Li, Xiaowen Su, and Di Jin. 2023. A Generalized Deep Markov Random Fields Framework for Fake News Detection. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI 2023, 19th-25th August 2023, Macao, SAR, China. ijcai.org, 4758– 4765

    work page 2023

  8. [8]

    Yixin Liu, Shiyuan Li, Yu Zheng, Qingfeng Chen, Chengqi Zhang, and Shirui Pan

Show all 61 references
  1. [9]

    Siqi Liu, Dongxiao He, Zhizhi Yu, Di Jin, and Zhiyong Feng. 2025. Beyond Homophily: Neighborhood Distribution-guided Graph Convolutional Networks. Expert Syst. Appl. 259 (2025), 125274

    2025

  2. [10]

    Junjun Pan, Yixin Liu, Yizhen Zheng, and Shirui Pan. 2023. PREM: A Simple Yet Effective Approach for Node-Level Graph Anomaly Detection. In 2023 IEEE International Conference on Data Mining (ICDM) . IEEE, 1253–1258

    2023

  3. [11]

    Xiaobao Wang, Yiqi Dong, Di Jin, Yawen Li, Longbiao Wang, and Jianwu Dang

  4. [12]

    Luzhi Wang, Yizhen Zheng, Di Jin, Fuyi Li, Yongliang Qiao, and Shirui Pan. 2024. Contrastive graph similarity networks. ACM Transactions on the Web 18, 2 (2024), 1–20

    2024

  5. [13]

    Wei Ju, Yifan Wang, Yifang Qin, Zhengyang Mao, Zhiping Xiao, Junyu Luo, Junwei Yang, Yiyang Gu, Dongjie Wang, Qingqing Long, Siyu Yi, Xiao Luo, and Ming Zhang. 2024. Towards Graph Contrastive Learning: A Survey and Beyond. CoRR abs/2405.11868 (2024). arXiv:2405.11868

    2024 arXiv

  6. [14]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Velickovic, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. 2019. Deep Graph Infomax. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net

    2019

  7. [15]

    Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Rémi Munos, Petar Veličković, and Michal Valko. 2021. Bootstrapped representation learning on graphs. In ICLR 2021 Workshop on Geometrical and Topological Representation Learning

    2021

  8. [16]

    Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S. Yu. 2021. From Canonical Correlation Analysis to Self-supervised Graph Neural Networks. In Advances in Neural Information Processing Systems, NeurIPS 2021 . 76–89

    2021

  9. [17]

    Jitao Zhao, Di Jin, Meng Ge, Lianze Shan, Xin Wang, Dongxiao He, and Zhiyong Feng. 2024. FUG: Feature-Universal Graph Contrastive Pre-training for Graphs with Diverse Node Features. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  10. [18]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. Deep Graph Contrastive Representation Learning. CoRR abs/2006.04131 (2020). arXiv:2006.04131

    2020 arXiv

  11. [19]

    Yizhen Zheng, Yu Zheng, Xiaofei Zhou, Chen Gong, Vincent CS Lee, and Shirui Pan. 2022. Unifying graph contrastive learning with flexible contextual scopes. In 2022 IEEE International Conference on Data Mining (ICDM) . IEEE, 793–802

    2022

  12. [20]

    Yifei Zhang, Hao Zhu, Zixing Song, Piotr Koniusz, and Irwin King. 2022. COSTA: Covariance-Preserving Feature Augmentation for Graph Contrastive Learning. In KDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022. ACM, 2524–2534

    2022

  13. [21]

    Mehdi Azabou, Venkataramana Ganesh, Shantanu Thakoor, Chi-Heng Lin, Lak- shmi Sathidevi, Ran Liu, Michal Valko, Petar Velickovic, and Eva L. Dyer. 2023. Half-Hop: A graph upsampling approach for slowing down message passing. In International Conference on Machine Learning, ICM...

    2023

  14. [22]

    Mengyue Liu, Yun Lin, Jun Liu, Bohao Liu, Qinghua Zheng, and Jin Song Dong

  15. [23]

    Dongxiao He, Lianze Shan, Jitao Zhao, Hengrui Zhang, Zhen Wang, and Weixiong Zhang. 2024. Exploitation of a Latent Mechanism in Graph Contrastive Learning: Representation Scattering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  16. [24]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph Contrastive Learning with Augmentations. In Advances in Neural Information Processing Systems, NeurIPS 2020, December 6-12, 2020, virtual

    2020

  17. [25]

    In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023

    B2-Sampling: Fusing Balanced and Biased Sampling for Graph Contrastive Learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023 . ACM, 1489–1500

    2023

  18. [26]

    Chunyu Wei, Yu Wang, Bing Bai, Kai Ni, David Brady, and Lu Fang. 2023. Boosting Graph Contrastive Learning via Graph Contrastive Saliency. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Proceedings of Machine Learning Resear...

    2023

  19. [27]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph Contrastive Learning with Adaptive Augmentation. In WWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021 . ACM / IW3C2, 2069–2080

    2021

  20. [28]

    Susheel Suresh, Pan Li, Cong Hao, and Jennifer Neville. 2021. Adversarial Graph Augmentation to Improve Graph Contrastive Learning. In Advances in Neural Information Processing Systems, NeurIPS 2021, December 6-14, 2021, virtual . 15920– 15933

    2021

  21. [29]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. CoRR abs/1807.03748 (2018). arXiv:1807.03748

    2018 arXiv

  22. [30]

    Dongxiao He, Jitao Zhao, Cuiying Huo, Yongqi Huang, Yuxiao Huang, and Zhiy- ong Feng. 2024. A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024. AAAI Press, 12340–12348

    2024

  23. [31]

    Jun Xia, Lirong Wu, Ge Wang, Jintao Chen, and Stan Z. Li. 2022. ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning. InInternational Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA (Proceedings of Machine Learning Research...

    2022

  24. [32]

    Namkyeong Lee, Junseok Lee, and Chanyoung Park. 2022. Augmentation-Free Self-Supervised Learning on Graphs. InThirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022. AAAI Press, 7372–7380

    2022

  25. [33]

    Wangbin Sun, Jintang Li, Liang Chen, Bingzhe Wu, Yatao Bian, and Zibin Zheng

  26. [34]

    Wen-Zhi Li, Chang-Dong Wang, Hui Xiong, and Jian-Huang Lai. 2023. HomoGCL: Rethinking Homophily in Graph Contrastive Learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023 . ACM, 1341–1352

    2023

  27. [35]

    Yixin Liu, Yizhen Zheng, Daokun Zhang, Vincent C. S. Lee, and Shirui Pan

  28. [36]

    Dongxiao He, Jitao Zhao, Rui Guo, Zhiyong Feng, Di Jin, Yuxiao Huang, Zhen Wang, and Weixiong Zhang. 2023. Contrastive Learning Meets Homophily: Two Birds with One Stone. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Procee...

    2023

  29. [37]

    In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, WSDM 2024, Merida, Mexico, March 4-8, 2024

    Rethinking and Simplifying Bootstrapped Graph Latents. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, WSDM 2024, Merida, Mexico, March 4-8, 2024 . ACM, 665–673

    2024

  30. [38]

    Teng Xiao, Huaisheng Zhu, Zhengyu Chen, and Suhang Wang. 2023. Simple and Asymmetric Graph Contrastive Learning without Augmentations. In Advances in Neural Information Processing Systems, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023

    2023

  31. [39]

    Zhouxia Wang, Tianshui Chen, Jimmy S. J. Ren, Weihao Yu, Hui Cheng, and Liang Lin. 2018. Deep Reasoning with Knowledge Graph for Social Relationship Understanding. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July ...

    2018

  32. [40]

    In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Washington, DC, USA, February 7-14, 2023

    Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Washington, DC, USA, February 7-14, 2023 . AAAI Press, 4516–4524

    2023

  33. [41]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of Graph Neural Network Evaluation. CoRR abs/1811.05868 (2018). arXiv:1811.05868

    2018 arXiv

  34. [42]

    Jiaming Zhuo, Can Cui, Kun Fu, Bingxin Niu, Dongxiao He, Chuan Wang, Yuan- fang Guo, Zhen Wang, Xiaochun Cao, and Liang Yang. 2024. Graph Contrastive Learning Reimagined: Exploring Universality. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, ...

    2024

  35. [43]

    Jingyu Chen, Runlin Lei, and Zhewei Wei. 2024. PolyGCL: GRAPH CON- TRASTIVE LEARNING via Learnable Spectral Polynomial Filters. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net

    2024

  36. [44]

    Jiaming Zhuo, Feiyang Qin, Can Cui, Kun Fu, Bingxin Niu, Mengzhu Wang, Yuan- fang Guo, Chuan Wang, Zhen Wang, Xiaochun Cao, and Liang Yang. 2024. Im- proving Graph Contrastive Learning via Adaptive Positive Sampling. InIEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2024

  37. [45]

    Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S. Yu. 2022. A Survey on Knowledge Graphs: Representation, Acquisition, and Applications. IEEE Trans. Neural Networks Learn. Syst. 33, 2 (2022), 494–514. WWW ’25, April 28–May 2, 2025, Sydney, NSW, Australia. ...

    2022

  38. [46]

    Yifei Zhang, Yankai Chen, Zixing Song, and Irwin King. 2023. Contrastive Cross-scale Graph Knowledge Synergy. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023. ACM, 3422–3433

    2023

  39. [47]

    Jonathon Shlens. 2014. A Tutorial on Principal Component Analysis. CoRR abs/1404.1100 (2014). arXiv:1404.1100

    2014 arXiv

  40. [48]

    Jitao Zhao, Dongxiao He, Meng Ge, Yongqi Huang, Lianze Shan, Yongbin Qin, and Zhiyong Feng. 2024. GA-GGD: Improving semantic discriminability in graph contrastive learning via Generative Adversarial Network. Inf. Fusion 110 (2024), 102465

    2024

  41. [49]

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. DeepWalk: online learning of social representations. In The 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’14, New York, NY, USA - August 24 - 27, 2014. ACM, 701–710

    2014

  42. [50]

    Harold Hotelling. 1992. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution . Springer, 162–190

    1992

  43. [51]

    Piotr Bielak, Tomasz Kajdanowicz, and Nitesh V. Chawla. 2022. Graph Barlow Twins: A self-supervised representation learning framework for graphs. Knowl. Based Syst. 256 (2022), 109631

    2022

  44. [52]

    Yizhen Zheng, Shirui Pan, Vincent C. S. Lee, Yu Zheng, and Philip S. Yu. 2022. Rethinking and Scaling Up Graph Contrastive Learning: An Extremely Efficient Approach with Group Discrimination. In Advances in Neural Information Process- ing Systems, NeurIPS 2022

    2022

  45. [53]

    Kaveh Hassani and Amir Hosein Khas Ahmadi. 2020. Contrastive Multi-View Representation Learning on Graphs. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020 (Proceedings of Machine Learning Research, Vol. 119). PMLR, 4116–4126

    2020

  46. [54]

    McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

    Julian J. McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

  47. [55]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for Networks. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17,

    2016

  48. [57]

    Hengrui Zhang, Qitian Wu, Yu Wang, Shaofeng Zhang, Junchi Yan, and Philip S. Yu. 2022. Localized Contrastive Learning on Graphs. CoRR abs/2212.04604 (2022). arXiv:2212.04604

    2022 arXiv

  49. [58]

    Cohen, and Ruslan Salakhutdinov

    Zhilin Yang, William W. Cohen, and Ruslan Salakhutdinov. 2016. Revisiting Semi-Supervised Learning with Graph Embeddings. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016 (JMLR Workshop and Conference Proceedings, Vol. 48). JMLR.org, 40–48

    2016

  50. [61]

    As outlined in the main text, we use the well-known GCL method, GRACE [18], and run it 20 times on each dataset under the default experimental settings

    Our goal is to identify nodes that are insufficiently trained, as evidenced by their frequent misclassification errors across multiple tests. As outlined in the main text, we use the well-known GCL method, GRACE [18], and run it 20 times on each dataset under the default exper...

    2025

  51. [2015]

    In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, Santiago, Chile, August 9-13, 2015

    Image-Based Recommendations on Styles and Substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, Santiago, Chile, August 9-13, 2015 . ACM, 43–52. Str-GCL: Structural Commonsense Driven Graph Contrastive ...

    2015

  52. [2023]

    In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Washington, DC, USA, February 7-14, 2023

    Augmenting Affective Dependency Graph via Iterative Incongruity Graph Learning for Sarcasm Detection. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Washington, DC, USA, February 7-14, 2023 . AAAI Press, 4702–4710

    2023

  53. [2024]

    CoRR abs/2405.16771 (2024)

    ARC: A Generalist Graph Anomaly Detector with In-Context Learning. CoRR abs/2405.16771 (2024). arXiv:2405.16771

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.