Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Mitigating Degree Bias Adaptively with Hard-to-Learn Nodes in Graph Contrastive Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A label-aware, hardness-adaptive contrastive loss reduces degree bias in graph node classification, and its semi-supervised extension keeps worst-case misclassification risk bounded.

desk verdict The empirical recipe is plausible and worth a look, but the theoretical appendix leans on a false sampling-ratio claim and should be treated as not established. read the letter →

arxiv 2506.05214 v1 pith:ZT6AGLEY submitted 2025-06-05 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords DegreeBiasGraphContrastiveLearningHardnessAdaptiveReweightedlossNodeClassificationSemi-supervisedPseudo-labelsStructuralFairness
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

Standard graph contrastive learning gives every node only one positive pair—its two augmented views—and weights all positives and negatives equally, which leaves low-degree nodes with thin and noisy supervision. The paper proposes the Hardness Adaptive Reweighted (HAR) loss, which uses node labels to create more positive pairs and then up-weights hard-to-learn positives and negatives, defined as same-origin augmented views and different-class nodes with high similarity. It further wraps HAR in a two-step semi-supervised pipeline, SHARP, that pre-trains on labelled nodes and fine-tunes on pseudo-labelled nodes. The intended result is that low-degree nodes receive more and cleaner contrastive signal, so their classification accuracy rises and the performance gap between low- and high-degree nodes narrows. Experiments on four citation and Wikipedia graphs with two GNN encoders report SHARP ahead of four baselines in 14 of 16 settings, with the largest gains in the lowest-degree groups.

What carries the argument

The load-bearing object is the HAR contrastive loss, a per-node objective of the form $\ell_i = -\log \frac{\mathrm{POS}_i}{\mathrm{POS}_i+\mathrm{NEG}_i}$, symmetrized over the two augmented views. POS entries are built from the label-based positive mask and a normalized similarity matrix $\bar S^+$ scaled by $\alpha$ plus an identity boost for same-origin pairs; NEG uses the negative mask, a class-balance ratio $Q_i$, and $\beta$-scaled negative similarities normalized by their mean. This design makes hard positives (same-origin) and hard negatives (similar but differently labelled) dominate the gradient, and because low-degree nodes appear disproportionately among hard negatives, the reweighting concentrates contrastive learning on them. The SHARP framework then extends HAR to semi-supervised graphs by pre-training on labelled nodes, generating pseudo-labels, and fine-tuning on the pseudo-labelled set, so the same adaptive weighting operates when labels are scarce.

What would settle it

Train SHARP on one of the four datasets, then compute the empirical ratio $w(x^-)=\sigma(g_p(x^-))/p(x^-)$ over the test nodes; if any input yields $w(x^-)<1$, the premise of Proposition C.2 fails and the corollary on bounded misclassification risk is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that the limiting factor for low-degree nodes in graph contrastive learning is not only the scarcity of positive pairs but the equal weighting of all contrastive pairs, and that replacing equal weighting with a hardness-adaptive, label-aware weighting reduces degree bias. For a node, HAR classifies other nodes as hard positives (same label and same augmentation origin), easy positives (same label, different origin), hard negatives (different label but high embedding similarity), and easy negatives (different label, low similarity); the loss then computes a debiasing-corrected ratio of reweighted positive and negative similarities and assigns higher weight to hard cases. Hard negatives are disproportionately low-degree in the empirical degree-distribution analysis, which is the mechanism the paper offers for why the weighting helps the underperforming group. The theoretical claim is that with $\alpha=0$ the HAR objective degenerates to the fully unsupervised contrastive case, so by the known worst-case bound for that case, SHARP's misclassification risk remains bounded even under worst-case hyperparameter settings. The four-dataset experiments comparing SHARP against GCN, GRACE, SCL, and Debias are presented as direct evidence for the degree-bias reduction.

Load-bearing premise

The proof assumes, without proving it, that the trained classifier's ratio of finding negative samples is never below the random-sampling rate, and that the discrete loss actually implemented equals the continuous expectation analysed; if either fails, the bounded-misclassification-risk corollary is not established, though the empirical degree-bias results would not necessarily fall.

Editorial extensions

If this is right

  • Low-degree node groups see the largest F1 improvements, with degree-1 nodes on Cora gaining about 11.9 percentage points over the GAT baseline at r=0.3, while high-degree groups change less.
  • SHARP beats the four baselines in 14 of 16 global test settings and reaches the highest gain on Cora (about 3.6-4.2 F1 points over the GNN baseline at r=0), while remaining competitive on PubMed.
  • Using only half the labelled data (r=0.5), SHARP with a GCN encoder reaches 0.8504 F1 on Cora, surpassing a plain GCN trained on all labels (0.834), indicating label efficiency.
  • Under worst-case hyperparameter settings ($\alpha=0$) the HAR loss is claimed to degenerate to the unsupervised contrastive objective, so the misclassification risk bound from that setting transfers to SHARP.
  • Performance varies by less than 2% across $\alpha$ and $\beta$ in (0,1], except when $\beta$ is near zero, where ignoring negatives degrades learning.

Reading between the lines

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

  • The hardness-weighting mechanism should transfer beyond the SHARP pipeline: applying HAR to other GCL variants, heterogeneous graphs, or inductive settings is a direct test of the claim, and the paper lists heterogeneous graphs as future work without testing it.
  • Because the bounded-risk guarantee passes through two unproven bridges—the assumption that the negative-sample probability ratio is never below 1 and the assumption that the discrete loss equals the continuous expectation analysed—a direct numerical check of both on the four datasets would decide whether the theoretical claim holds as stated.
  • If labels contain noise, the hard-negative definition (similar but differently labelled) will treat mislabelled positives as hard negatives; HAR's sensitivity to label noise is untested, and the reported drop below 30% labelled data is consistent with such fragility.
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 / 6 minor

Summary. This paper proposes a Hardness Adaptive Reweighted (HAR) contrastive loss for graph node classification, which uses label information to define positives and negatives and adaptively reweights pairs according to learning hardness. The HAR loss is then wrapped in a two-step semi-supervised framework called SHARP, where a model is pre-trained on labelled data and fine-tuned on pseudo-labelled data. The main empirical claim is that SHARP improves global F1 and low-degree node performance across Cora, CiteSeer, PubMed, and Wiki-CS with both GCN and GAT encoders. The paper also claims a theoretical guarantee that, even under worst-case hyperparameter settings, SHARP's misclassification risk remains bounded, based on adapting the analysis of Robinson et al. (2021).

Significance. If the empirical claims hold, the HAR loss is a simple, practical modification to graph contrastive learning that could improve structural fairness without sacrificing global accuracy, and the semi-supervised extension is plausibly useful in label-scarce settings. The paper includes extensive experiments across four datasets and two encoders, degree-level breakdowns, t-SNE visualisations, and sensitivity analysis, and it explicitly reports the number of runs. However, the theoretical analysis currently rests on an unproven and in general false sampling-ratio assertion and on an unverified identification between the implemented discrete loss and the continuous objective that is actually analysed; these gaps are load-bearing for the theory but not for the core empirical contribution.

major comments (3)
  1. [Appendix C, Proposition C.2] The proposition asserts w(x^-)=sigma_sigmoid(gp(x^-))/p(x^-) >= 1 for every negative sample, with no proof and only an intuitive justification, and no calibration or domination assumption is stated. In general the assertion is false: for p(x^-)=0.8 and sigma_sigmoid(gp(x^-))=0.5, w(x^-)=0.625<1. The proof of Lemma C.1 uses this inequality twice, both to bound the bad-event probability by E[w] and to lower-bound Z_beta by removing w from the integral. If w(x^-)<1 on a set of positive measure, the e^{-beta*epsilon/2} decay argument collapses, so Corollary C.1 is not established by the material in the appendix.
  2. [Section 3.2 and Appendix C, Eq. (16)] The loss whose bound is proved is not shown to be the loss that is implemented. Equations (10)-(11) use mask matrices Mask+ and Mask- built from ground-truth labels and reweighted similarities W+ from Eqs. (6)-(7), including the min-max normalization and the identity-matrix boost for hard positives; Eq. (16) instead contains alpha*e^{-g_bar(x)g_bar(x+)} for non-augmentation positives and a sigmoid classifier sigma_sigmoid(gp(x-)) that does not appear in the HAR loss definition. No derivation links these two forms, so the boundedness result may apply to a different objective than the one evaluated in the experiments.
  3. [Section 4.2, Tables 3, 6, 8] The global claim that SHARP outperforms baselines in 14 of 16 tests is reported without error bars or significance tests, even though each model is run 10 times; means alone do not establish that the improvements are not noise. More importantly, the blanket statement that SHARP improves at the degree level is contradicted by Table 8, where for degree-6 nodes on Cora with GAT and r=0 SHARP (0.7846) is 6.16% below the GAT baseline (0.8462), and by Table 6, where SHARP is also below GCN at degree 6 (0.7865 vs 0.8077). The degree-bias claim therefore needs qualification and, ideally, per-group standard deviations.
minor comments (6)
  1. [Figure 5] The gray area is described as an ablation study, but it is not clear whether this is the degree distribution under uniform weighting or the degree distribution of all nodes; please label the curves and define the baseline distribution explicitly.
  2. [Table 3] The precision of reported values is inconsistent (e.g., '0.726' with two decimals versus three decimals elsewhere); report all values with the same number of significant digits.
  3. [Appendix C] The phrase 'with slight abuse of notations' and the remark that increasing alpha reduces positive weights should be made precise, since Eq. (6) has alpha as a positive scaling factor of the normalized similarity, not of a negative exponential.
  4. [References] The citation to Velickovic et al. (2017) contains corrupted characters in the body text; please fix the citation encoding.
  5. [Section 3.2] The definition of hard positives as same-origin augmented pairs is not explicitly connected to the weighting formulas; the identity-matrix term in Eq. (6) implements this, but the text should state the correspondence.
  6. [Section 4.2.3] The reported decline in performance when beta is close to 0 is not visible in the panels shown; include the corresponding subfigure or a table with the beta sweep values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SHARP's empirical claims are self-contained, and the theoretical bound is imported from external prior work rather than derived from the paper's own fitted quantities.

full rationale

Walking the claimed derivation chain, the central empirical claim (SHARP improves global and degree-level F1) is supported by held-out test evaluations on four datasets, with hyperparameters α and β examined by sensitivity sweeps rather than fitted to the reported metric. The hardness weighting is computed from the model's own similarities, which is self-referential training dynamics but not a definitional equivalence between input and output. The theoretical bound is explicitly carri ed over from Robinson et al. (2021): the paper states that at α = 0 the loss degenerates to that prior unsupervised objective and then imports the corresponding lemmas. That is external evidence, not a self-citation chain. The appendix's Proposition C.2 is asserted without proof ('Intuitively...') and the identification of the discrete HAR loss with the continuous expectation in Eq. (16) is not formally derived; these are correctness or rigor gaps that lower confidence in the theoretical validation, but they do not reduce the result to the paper's own inputs by construction. No relevant self-citation, uniqueness import, ansatz smuggling, or renaming of a known result as a new finding is present. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The ledger shows that the central empirical claim rests on standard GCL assumptions plus a heuristic weighting scheme. The theoretical claim rests on two unproven assumptions about the classifier and the representation of the loss.

free parameters (5)
  • alpha (positive reweighting scale) = in (0,1], sensitivity analysis on Cora shows robust range
    Scale for easy positives in Equation (6); tuned by sensitivity sweep, not derived.
  • beta (negative reweighting scale) = in (0,1], sensitivity analysis on Cora
    Scale for hard negatives in Equation (9); chosen by hand, affects loss level.
  • temperature tau = 0.4 (Cora), 0.9 (CiteSeer), 0.7 (PubMed), 0.4 (Wiki-CS)
    Appears in f_tau similarity and loss bound; per-dataset tuning.
  • edge/feature drop rates pe, pf = not specified per dataset, only range [0,0.4]
    Augmentation strength in Section 3.1; values not reported, so replication requires guessing.
  • unlabelled ratio r = 0 and 0.3 in experiments; r=0.5 mentioned in text
    Controls label/unlabelled split in SHARP; experimental variable.
assumptions (5)
  • domain assumption The GCL framework with random masking of edges and features produces useful augmented views.
    Standard assumption in graph contrastive learning, used throughout Section 3.1.
  • domain assumption Adding more positives from same-class nodes and weighting by hardness reduces degree bias.
    This is the core intuition of the paper, not proven independently; validated only by experiments.
  • ad hoc to paper The sigmoid classifier's negative probability ratio w(x-) is at least 1 for all x- (Proposition C.2).
    Asserted in Appendix C without proof; needed for the bound to hold.
  • ad hoc to paper The actual HAR loss (Equations 10-11) is equivalent to the expectation L(g,alpha,beta) in Equation (16).
    The paper states 'Given the symmetric property... the loss function can be expressed as the expectation' but no formal derivation is given.
  • domain assumption Pseudo-labels from the pretrained model are reliable enough for fine-tuning.
    Assumed in SHARP's two-step framework; the paper notes limitations when labelled data is scarce.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Degree Bias Adaptively with Hard-to-Learn Nodes in Graph Contrastive Learning." pith.science (2026). https://pith.science/paper/ZT6AGLEY

@misc{pith2026250605214,
  author       = {Pith},
  title        = {Pith review of: Mitigating Degree Bias Adaptively with Hard-to-Learn Nodes in Graph Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZT6AGLEY}},
  note         = {Machine review of arXiv:2506.05214}
}
read the original abstract

Graph Neural Networks (GNNs) often suffer from degree bias in node classification tasks, where prediction performance varies across nodes with different degrees. Several approaches, which adopt Graph Contrastive Learning (GCL), have been proposed to mitigate this bias. However, the limited number of positive pairs and the equal weighting of all positives and negatives in GCL still lead to low-degree nodes acquiring insufficient and noisy information. This paper proposes the Hardness Adaptive Reweighted (HAR) contrastive loss to mitigate degree bias. It adds more positive pairs by leveraging node labels and adaptively weights positive and negative pairs based on their learning hardness. In addition, we develop an experimental framework named SHARP to extend HAR to a broader range of scenarios. Both our theoretical analysis and experiments validate the effectiveness of SHARP. The experimental results across four datasets show that SHARP achieves better performance against baselines at both global and degree levels.

Figures

Figures reproduced from arXiv: 2506.05214 by the authors.

Figure 1
Figure 1. (Left.) The degree bias example in node classification: GNNs can perform worse on low-degree nodes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The Workflow of Contrastive Learning variations (Chen et al., 2020; Assran et al., 2020; Khosla et al., 2020), most GCL methods still use the common contrastive loss: it treats two aug￾mented node originated from the same node as the positive pair, while treating all others as equally im￾portant negatives. This can make it hard to handle degree bias as we will discuss in detail below. 2.2 Degree Bias in Node Classif… view at source ↗
Figure 3
Figure 3. The Illustration of Graph Contrastive Learning with Our Proposed HAR Loss [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The Two-Step SHARP with HAR Loss on Node Classification Tasks [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The Node Degree Distributions of Top-5 Hard [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: (a) Degree Bias Comparison; (b)(c) r, α, β Sensitivity Analysis on Cora Dataset than 2% across different parameter combinations. However, when β is particularly small (close to 0), a noticeable decline occurs. This implies that ignor￾ing negative samples can cause the …
Figure 7
Figure 7. Figure 7: The t-SNE Visualizations of Embeddings on Cora and CiteSeer Dataset [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: The t-SNE visualizations of embeddings on different models (GAT Layer, [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 26 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Mahmoud Assran, Nicolas Ballas, Lluis Castrejon, and Michael Rabbat. 2020. Supervision accelerates pre-training in contrastive semi-supervised learning of visual representations. arXiv preprint arXiv:2006.10803

  4. [4]

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veli c kovi \'c . 2021. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PMLR

  6. [6]

    Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. 2020. Debiased contrastive learning. Advances in neural information processing systems, 33:8765--8775

  7. [7]

    Fuli Feng, Weiran Huang, Xiangnan He, Xin Xin, Qifan Wang, and Tat-Seng Chua. 2021. Should graph convolution trust neighbors? a simple causal inference method. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1208--1218

  8. [8]

    Matthias Fey and Jan E. Lenssen. 2019. Fast graph representation learning with PyTorch Geometric . In ICLR Workshop on Representation Learning on Graphs and Manifolds

Show all 41 references
  1. [9]

    C Lee Giles, Kurt D Bollacker, and Steve Lawrence. 1998. Citeseer: An automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries, pages 89--98

  2. [10]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems, 30

  3. [11]

    Xiaotian Han, Kaixiong Zhou, Ting-Hsiang Wang, Jundong Li, Fei Wang, and Na Zou. 2024. Marginal nodes matter: Towards structure fairness in graphs. ACM SIGKDD Explorations Newsletter, 25(2):4--13

  4. [12]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729--9738

  5. [13]

    Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. 2020. Hard negative mixing for contrastive learning. Advances in neural information processing systems, 33:21798--21809

  6. [14]

    Jian Kang and Hanghang Tong. 2021. Fair graph mining. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 4849--4852

  7. [15]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning. Advances in neural information processing systems, 33:18661--18673

  8. [16]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

  9. [17]

    Oyku Deniz Kose and Yanning Shen. 2022. Fair contrastive learning on graphs. IEEE Transactions on Signal and Information Processing over Networks, 8:475--488

  10. [18]

    Junnan Li, Caiming Xiong, and Steven CH Hoi. 2021. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475--9484

  11. [19]

    Langzhang Liang, Zenglin Xu, Zixing Song, Irwin King, Yuan Qi, and Jieping Ye. 2023. Tackling long-tailed distribution issue in graph neural networks via normalization. IEEE Transactions on Knowledge and Data Engineering

  12. [20]

    Hongyi Ling, Zhimeng Jiang, Youzhi Luo, Shuiwang Ji, and Na Zou. 2023. Learning fair graph representations via automated data augmentations. In International Conference on Learning Representations (ICLR)

  13. [21]

    Zemin Liu, Trung-Kien Nguyen, and Yuan Fang. 2021. Tail-gnn: Tail-node graph neural networks. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1109--1119

  14. [22]

    Zemin Liu, Trung-Kien Nguyen, and Yuan Fang. 2023. On generalized degree fairness in graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 4525--4533

  15. [23]

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. 2000. Automating the construction of internet portals with machine learning. Information Retrieval, 3:127--163

  16. [24]

    P \'e ter Mernyei and C a t a lina Cangea. 2020. Wiki-cs: A wikipedia-based benchmark for graph neural networks. arXiv preprint arXiv:2007.02901

  17. [25]

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

  18. [26]

    Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. 2021. Contrastive learning with hard negative samples. International Conference on Learning Representations (ICLR)

  19. [27]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data. AI magazine, 29(3):93--93

  20. [28]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929--1958

  21. [29]

    Arjun Subramonian, Jian Kang, and Yizhou Sun. 2024. Theoretical and empirical insights into the origins of degree bias in graph neural networks. arXiv preprint arXiv:2404.03139

  22. [30]

    Xianfeng Tang, Huaxiu Yao, Yiwei Sun, Yiqi Wang, Jiliang Tang, Charu Aggarwal, Prasenjit Mitra, and Suhang Wang. 2020. Investigating and mitigating degree-related biases in graph convoltuional networks. In Proceedings of the 29th ACM International Conference on Information & K...

  23. [31]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11)

  24. [32]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2017. Graph attention networks. 6th International Conference on Learning Representations

  25. [33]

    Ruijia Wang, Xiao Wang, Chuan Shi, and Le Song. 2022. Uncovering the structural fairness in graph contrastive learning. Advances in neural information processing systems, 35:32465--32473

  26. [34]

    Teng Xiao, Zhengyu Chen, Donglin Wang, and Suhang Wang. 2021. Learning how to propagate messages in graph neural networks. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1894--1903

  27. [35]

    Hao Yan, Chaozhuo Li, Ruosong Long, Chao Yan, Jianan Zhao, Wenwen Zhuang, Jun Yin, Peiyan Zhang, Weihao Han, Hao Sun, et al. 2023. A comprehensive study on text-attributed graphs: Benchmarking and rethinking. Advances in Neural Information Processing Systems, 36:17238--17264

  28. [36]

    Sukwon Yun, Kibum Kim, Kanghoon Yoon, and Chanyoung Park. 2022. Lte4g: Long-tail experts for graph neural networks. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 2434--2443

  29. [37]

    Shengzhong Zhang, Wenjie Yang, Yimin Zhang, Hongwei Zhang, Divin Yan, and Zengfeng Huang. 2023. Understanding community bias amplification in graph representation learning. arXiv preprint arXiv:2312.04883

  30. [38]

    Mingxia Zhao and Adele Lu Jia. 2024. Dahgn: Degree-aware heterogeneous graph neural network. Knowledge-Based Systems, 285:111355

  31. [39]

    Shuai Zhao, Tianyu Zhang, Man Hu, Wen Chang, and Fucheng You. 2022. Ap-bert: enhanced pre-trained model through average pooling. Applied Intelligence, 52(14):15929--15937

  32. [40]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2020. http://arxiv.org/abs/2006.04131 Deep graph contrastive representation learning . In ICML Workshop on Graph Representation Learning and Beyond

  33. [41]

    Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph contrastive learning with adaptive augmentation. In Proceedings of the web conference 2021, pages 2069--2080

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.