pith. sign in

arxiv: 2605.03076 · v1 · submitted 2026-05-04 · 💻 cs.LG

Adaptive Negative Scheduling for Graph Contrastive Learning

Adnan Ali , Jinlong Li , Syed Muhammad Israr , Ali Kashif Bashir This is my paper

Pith reviewed 2026-05-08 18:37 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph contrastive learningnegative samplingadaptive schedulinghardness-aware selectionself-supervised learningrepresentation learningcomputational budget
0
0 comments X

The pith

AdNGCL adaptively schedules negatives in graph contrastive learning by tracking loss trends under per-category budgets.

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

The paper introduces AdNGCL, a framework that replaces fixed negative sampling in graph contrastive learning with a dynamic scheduler. The scheduler classifies negatives into hard, intermediate, and easy groups and adjusts how many of each to draw according to current contrastive loss behavior while staying inside global and category-specific compute budgets. It also refreshes the pool of candidates at intervals to keep variety. If this works, representations learned without labels should become more accurate on graph tasks such as recommendation and anomaly detection, and practitioners gain direct knobs for trading accuracy against training time. Experiments across nine standard graph datasets show the method reaches the highest accuracy on seven of them and the second-highest on the other two.

Core claim

Negative selection in graph contrastive learning can be cast as a loss-gated, budget-constrained allocation across hardness strata. The HANS scheduler monitors contrastive loss trends and raises or lowers the step size for each stratum so that the training process spends compute where the loss signal indicates the greatest remaining gain, while a periodic refresh step keeps the candidate pool diverse inside the same budgets. This produces the reported accuracy gains on the nine datasets together with explicit control over total negative-sample cost.

What carries the argument

The hardness-aware scheduler (HANS) that formulates negative selection as a loss-gated, budget-constrained process across hard, intermediate, and easy strata and dynamically adjusts step sizes based on contrastive loss trends.

If this is right

  • The method reaches highest accuracy on seven of nine standard graph datasets and second-highest on the remaining two.
  • It supplies explicit global and per-category budgets that let users cap the number of negatives processed per epoch.
  • Periodic refresh of the candidate pool maintains diversity inside the same budgets.
  • The same loss-trend logic can be applied to any contrastive objective that already stratifies negatives by hardness.
  • Representation quality improves while total negative-sample computation stays bounded.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same budget-aware logic could be ported to non-graph contrastive settings such as images or sequences where negative sampling is also expensive.
  • Per-category budgets might be tuned to graph properties such as degree distribution or community structure.
  • If the scheduler reduces reliance on the very hardest negatives early in training, it could lessen gradient instability that sometimes appears in contrastive objectives.
  • An open question left by the work is whether the observed gains remain when the underlying encoder architecture changes.

Load-bearing premise

Dynamically adjusting step sizes based on contrastive loss trends under global and per-category budgets will reliably improve representation quality without introducing training instability or selection bias.

What would settle it

A controlled comparison on the same nine datasets in which the adaptive step-size rule is replaced by fixed or randomly jittered step sizes, showing no accuracy lift or an increase in run-to-run variance.

Figures

Figures reproduced from arXiv: 2605.03076 by Adnan Ali, Ali Kashif Bashir, Jinlong Li, Syed Muhammad Israr.

Figure 1
Figure 1. Figure 1: Schematic diagram of the AdNGCL framework. It illustrates the full training pipeline, including data augmentation, contrastive learning, adaptive view at source ↗
Figure 2
Figure 2. Figure 2: A structural diagram of Hardness-Aware Negative Scheduling view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualizations of node embeddings learned by AdNGCL view at source ↗
Figure 4
Figure 4. Figure 4: Temporal analysis of HANS. Each subplot displays the epoch-wise view at source ↗
Figure 5
Figure 5. Figure 5: Average training time (ms) over 2000 epochs on WikiCS at 50% (left) view at source ↗
read the original abstract

Graph contrastive learning (GCL) has become a central paradigm for self-supervised representation learning in computational intelligence, with applications spanning recommendation, anomaly detection, and personalization. A key limitation of existing methods is their reliance on static negative sampling, which fails to account for the dynamic informativeness and computational cost of negatives during training. We propose AdNGCL, an adaptive negative scheduling framework with a hardness-aware scheduler (HANS) that formulates negative selection as a loss-gated, budget-constrained process across hard, intermediate, and easy strata. The scheduler dynamically adjusts step sizes based on contrastive loss trends under both global and per-category budgets, while periodically refreshing samples to maintain diversity without exceeding compute constraints. Experiments on nine benchmark graph datasets demonstrate that AdNGCL consistently advances state-of-the-art performance, achieving the best accuracy on seven datasets and second-best on the remaining two, while offering explicit control over computational cost. These results highlight the value of budget-aware, loss-sensitive scheduling as a general strategy for improving the robustness and efficiency of representation learning in emerging computational intelligence applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes AdNGCL, a framework for graph contrastive learning that replaces static negative sampling with an adaptive hardness-aware negative scheduler (HANS). HANS treats negative selection as a loss-gated, budget-constrained process over hard/intermediate/easy strata, dynamically adjusting step sizes from contrastive loss trends under global and per-category budgets while periodically refreshing samples. Experiments on nine benchmark graph datasets are reported to show consistent gains, with best accuracy on seven datasets and second-best on two, plus explicit control over computational cost.

Significance. If the empirical results hold under proper controls and the scheduler can be shown to be the load-bearing component rather than periodic refresh or budget allocation alone, the work would offer a practical, budget-aware strategy for improving efficiency and robustness in GCL. The explicit handling of computational constraints and stratification is a useful direction for self-supervised graph methods.

major comments (3)
  1. [Method / HANS scheduler description] The manuscript provides no derivation, pseudocode, or stability analysis for the HANS loss-gated step-size adjustment rule. Without this, it is impossible to determine whether the adaptive component improves representation quality or merely introduces oscillation or stratum bias, as the central claim requires.
  2. [Experiments] Experiments section: No error bars, standard deviations, or statistical significance tests are reported for the accuracy gains across the nine datasets. The headline claim of consistent SOTA advances therefore cannot be verified from the presented results.
  3. [Experiments / Ablation studies] Experiments section: No ablation studies isolate the contribution of the adaptive loss-based adjustment from the periodic refresh mechanism or the global/per-category budget allocation. The results could be driven by the latter two components alone, undermining the claim that the scheduler itself is responsible for the observed improvements.
minor comments (2)
  1. [Method] Notation for the strata (hard, intermediate, easy) and the exact definitions of global vs. per-category budgets should be introduced with equations or a clear table early in the method section.
  2. [Experiments] The abstract states 'advances state-of-the-art performance' but the full results table should explicitly list the prior SOTA baselines for each dataset to allow direct comparison.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to strengthen the description of HANS, the experimental reporting, and the ablation analysis.

read point-by-point responses

  1. Referee: [Method / HANS scheduler description] The manuscript provides no derivation, pseudocode, or stability analysis for the HANS loss-gated step-size adjustment rule. Without this, it is impossible to determine whether the adaptive component improves representation quality or merely introduces oscillation or stratum bias, as the central claim requires.

    Authors: We agree that the current manuscript would benefit from greater formalization of the HANS scheduler. In the revised version we will add a derivation of the loss-gated step-size rule based on observed contrastive loss trends, include complete pseudocode for the full scheduling procedure (including global and per-category budget constraints), and provide an empirical stability analysis tracking step-size trajectories and stratum allocations across epochs to show that adaptation improves representation quality without inducing oscillation or systematic bias. revision: yes

  2. Referee: [Experiments] Experiments section: No error bars, standard deviations, or statistical significance tests are reported for the accuracy gains across the nine datasets. The headline claim of consistent SOTA advances therefore cannot be verified from the presented results.

    Authors: We acknowledge the absence of variability measures and statistical tests. We will rerun all experiments with multiple random seeds, report mean accuracies together with standard deviations, and include statistical significance tests (e.g., paired t-tests or Wilcoxon signed-rank tests) comparing AdNGCL against the strongest baselines on each dataset. revision: yes

  3. Referee: [Experiments / Ablation studies] Experiments section: No ablation studies isolate the contribution of the adaptive loss-based adjustment from the periodic refresh mechanism or the global/per-category budget allocation. The results could be driven by the latter two components alone, undermining the claim that the scheduler itself is responsible for the observed improvements.

    Authors: We recognize the value of isolating the adaptive component. The revised manuscript will contain new ablation experiments that (i) disable the loss-gated step-size adjustment while retaining periodic refresh and budget allocation, (ii) remove periodic refresh while keeping adaptation, and (iii) replace per-category budgets with uniform allocation. These controlled variants will quantify the specific contribution of the adaptive loss-based rule. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims rest on external benchmarks without self-referential reductions

full rationale

The abstract and description present AdNGCL as a scheduling framework whose core is a loss-gated dynamic step-size adjustment under budgets, with performance validated on nine independent graph datasets. No equations, parameter-fitting procedures, or self-citations appear in the supplied text that would allow any claimed prediction or result to reduce by construction to the method's own inputs. The evaluation metrics (accuracy on held-out benchmarks) are external and falsifiable, satisfying the criteria for a self-contained, non-circular derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The framework rests on the assumption that negatives can be meaningfully stratified by hardness and that loss trends provide a reliable signal for budget allocation; no explicit free parameters or invented entities are quantified in the abstract.

free parameters (2)
  • global and per-category budgets
    Used to constrain negative selection across strata; values not specified in abstract.
  • step sizes
    Dynamically adjusted based on loss trends; chosen during training.
axioms (2)
  • domain assumption Negative samples can be stratified into hard, intermediate, and easy categories based on informativeness
    Invoked in the design of the HANS scheduler.
  • domain assumption Contrastive loss trends reliably indicate when to refresh or adjust negative samples
    Central to the loss-gated scheduling logic.
invented entities (1)
  • HANS scheduler no independent evidence
    purpose: Formulates negative selection as a loss-gated, budget-constrained process
    New component introduced by the paper

pith-pipeline@v0.9.0 · 5487 in / 1391 out tokens · 56052 ms · 2026-05-08T18:37:48.319678+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    Graph contrastive learning for multi-behavior recommender system,

    Z. Zhao, J. Liu, N. Huang, P. Jiao, and H. Wu, “Graph contrastive learning for multi-behavior recommender system,”IEEE Transactions on Consumer Electronics, vol. 71, no. 2, pp. 7269–7279, May 2025

    work page 2025

  2. [2]

    Unsupervised anomaly detection on attributed networks with graph contrastive learning for consumer electronics security,

    B. Xu, J. Wang, Z. Zhao, H. Lin, and F. Xia, “Unsupervised anomaly detection on attributed networks with graph contrastive learning for consumer electronics security,”IEEE Transactions on Consumer Elec- tronics, vol. 70, no. 1, pp. 4062–4072, Feb 2024

    work page 2024

  3. [3]

    Incremental updates based on graph theory for consumer electronic devices,

    G. Ni, Z. Chen, J. Jiang, J. Luo, and Y . Ma, “Incremental updates based on graph theory for consumer electronic devices,”IEEE Transactions on Consumer Electronics, vol. 61, no. 1, pp. 128–136, February 2015

    work page 2015

  4. [4]

    Graph contrastive learning with feature augmentation for rumor detection,

    S. Li, W. Li, A. M. Luvembe, and W. Tong, “Graph contrastive learning with feature augmentation for rumor detection,”IEEE Transactions on Computational Social Systems, vol. 11, no. 4, pp. 5158–5167, 2023

    work page 2023

  5. [5]

    Affinity uncertainty-based hard negative mining in graph contrastive learning,

    C. Niu, G. Pang, and L. Chen, “Affinity uncertainty-based hard negative mining in graph contrastive learning,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 9, pp. 11 681–11 691, 2024

    work page 2024

  6. [6]

    A comprehensive survey on deep graph representation learning,

    W. Ju, Z. Fang, Y . Gu, Z. Liu, Q. Long, Z. Qiao, Y . Qin, J. Shen, F. Sun, Z. Xiao, J. Yang, J. Yuan, Y . Zhao, Y . Wang, X. Luo, and M. Zhang, “A comprehensive survey on deep graph representation learning,”Neural Networks, vol. 173, p. 106207, 2024

    work page 2024

  7. [7]

    Graph-based semi-supervised learning: A comprehensive review,

    Z. Song, X. Yang, Z. Xu, and I. King, “Graph-based semi-supervised learning: A comprehensive review,”IEEE Transactions on Neural Net- works and Learning Systems, vol. 34, no. 11, pp. 8174–8194, 2023. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, MM 202X 13

    work page 2023

  8. [8]

    Graph barlow twins: A self-supervised representation learning framework for graphs,

    P. Bielak, T. Kajdanowicz, and N. V . Chawla, “Graph barlow twins: A self-supervised representation learning framework for graphs,” Knowledge-Based Systems, vol. 256, p. 109631, 2022

    work page 2022

  9. [9]

    A simple frame- work for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple frame- work for contrastive learning of visual representations,” inProceedings of the 37th International Conference on Machine Learning (ICML). JMLR.org, 2020

    work page 2020

  10. [10]

    Multi-relational graph contrastive learning with learnable graph augmentation,

    X. Mo, J. Pang, B. Wan, R. Tang, H. Liu, and S. Jiang, “Multi-relational graph contrastive learning with learnable graph augmentation,”Neural Networks, vol. 181, p. 106757, 2025

    work page 2025

  11. [11]

    Graph contrastive learning with augmentations,

    Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen, “Graph contrastive learning with augmentations,” inProceedings of the 34th International Conference on Neural Information Processing Systems (NeurIPS). Curran Associates Inc., 2020

    work page 2020

  12. [12]

    Negative samples selecting strategy for graph contrastive learning,

    R. Miao, Y . Yang, Y . Ma, X. Juan, H. Xue, J. Tang, Y . Wang, and X. Wang, “Negative samples selecting strategy for graph contrastive learning,”Information Sciences, vol. 613, pp. 667–681, 2022

    work page 2022

  13. [13]

    Graph contrastive learning with implicit augmentations,

    H. Liang, X. Du, B. Zhu, Z. Ma, K. Chen, and J. Gao, “Graph contrastive learning with implicit augmentations,”Neural Networks, vol. 163, pp. 156–164, 2023

    work page 2023

  14. [14]

    Graph self-supervised learning: A survey,

    Y . Liu, M. Jin, S. Pan, C. Zhou, Y . Zheng, F. Xia, and P. S. Yu, “Graph self-supervised learning: A survey,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 6, pp. 5879–5900, 2023

    work page 2023

  15. [15]

    Graph data augmentation for graph machine learning: A survey,

    T. Zhao, W. Jin, Y . Liu, Y . Wang, G. Liu, S. G ¨unnemann, N. Shah, and M. Jiang, “Graph data augmentation for graph machine learning: A survey,”IEEE Data Engineering Bulletin, vol. 46, no. 2, pp. 140–165, 2023

    work page 2023

  16. [16]

    Graph contrastive learning with adaptive augmentation,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Graph contrastive learning with adaptive augmentation,” inThe Web Conference 2021 (WWW). ACM, 2021, pp. 2069–2080

    work page 2021

  17. [17]

    Features based adaptive augmentation for graph contrastive learning,

    A. Ali and J. Li, “Features based adaptive augmentation for graph contrastive learning,”Digital Signal Processing, vol. 145, p. 104312, 2024. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S1051200423004074

    work page 2024

  18. [18]

    A multi-view graph contrastive learning framework for defending against adversarial attacks,

    F. Cao, X. Ye, and H. Ye, “A multi-view graph contrastive learning framework for defending against adversarial attacks,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 8, no. 6, pp. 4022–4032, Dec 2024

    work page 2024

  19. [19]

    Adversarial graph augmentation to improve graph contrastive learning,

    S. Suresh, P. Li, C. Hao, and J. Neville, “Adversarial graph augmentation to improve graph contrastive learning,” inProceedings of the 35th International Conference on Neural Information Processing Systems (NeurIPS). Curran Associates Inc., 2021

    work page 2021

  20. [20]

    Adaptive data augmentation for con- trastive learning,

    Y . Zhang, H. Zhu, and S. Yu, “Adaptive data augmentation for con- trastive learning,” in2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, June 2023, pp. 1–5

    work page 2023

  21. [21]

    Deep graph contrastive representation learning,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Deep graph contrastive representation learning,” 2020

    work page 2020

  22. [22]

    CuCo: Graph representation with curriculum contrastive learning,

    G. Chu, X. Wang, C. Shi, and X. Jiang, “CuCo: Graph representation with curriculum contrastive learning,” inProceedings of the 13th In- ternational Joint Conference on Artificial Intelligence (IJCAI). IJCAI Organization, 8 2021, pp. 2300–2306, main Track

    work page 2021

  23. [23]

    Contrastive learning with hard negative samples,

    J. D. Robinson, C.-Y . Chuang, S. Sra, and S. Jegelka, “Contrastive learning with hard negative samples,” inInternational Conference on Learning Representations (ICLR), 2021

    work page 2021

  24. [24]

    From overfitting to robustness: Quantity, quality, and variety oriented negative sample selection in graph contrastive learning,

    A. Ali, J. Li, H. Chen, and A. K. Bashir, “From overfitting to robustness: Quantity, quality, and variety oriented negative sample selection in graph contrastive learning,”Applied Soft Computing, vol. 170, p. 112672, 2025. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S1568494624014467

    work page 2025

  25. [25]

    Generating counterfactual hard negative samples for graph contrastive learning,

    H. Yang, H. Chen, S. Zhang, X. Sun, Q. Li, X. Zhao, and G. Xu, “Generating counterfactual hard negative samples for graph contrastive learning,” inProceedings of the ACM Web Conference 2023 (WWW). ACM, 2023, pp. 621–629

    work page 2023

  26. [26]

    Boosting graph contrastive learning via adaptive sampling,

    S. Wan, Y . Zhan, S. Chen, S. Pan, J. Yang, D. Tao, and C. Gong, “Boosting graph contrastive learning via adaptive sampling,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 11, pp. 15 971–15 983, Nov 2023

    work page 2023

  27. [27]

    Deep adversarial metric learning,

    Y . Duan, W. Zheng, X. Lin, J. Lu, and J. Zhou, “Deep adversarial metric learning,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, June 2018, pp. 2780–2789

    work page 2018

  28. [28]

    ProGCL: Rethinking hard negative mining in graph contrastive learning,

    J. Xia, L. Wu, G. Wang, J. Chen, and S. Z. Li, “ProGCL: Rethinking hard negative mining in graph contrastive learning,” inProceedings of the 39th International Conference on Machine Learning (ICML). PMLR, 17-23 Jul 2022, pp. 24 332–24 346

    work page 2022

  29. [29]

    Adversarial contrastive graph augmentation with counterfactual regularization,

    T. Long, L. Zhang, L. Zhang, and L. Cui, “Adversarial contrastive graph augmentation with counterfactual regularization,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 18, pp. 19 086–19 094, Apr. 2025. [Online]. Available: https://ojs.aaai.org/ index.php/AAAI/article/view/34101

    work page 2025

  30. [30]

    Hard sample aware network for contrastive deep graph clustering,

    Y . Liu, X. Yang, S. Zhou, X. Liu, Z. Wang, K. Liang, W. Tu, L. Li, J. Duan, and C. Chen, “Hard sample aware network for contrastive deep graph clustering,”Proceedings of the AAAI Conference on Artificial Intelligence, no. 7, pp. 8914–8922, Jun. 2023

    work page 2023

  31. [31]

    Fir-gnn: A graph neural network using flow interaction relationships for intrusion detection of consumer electronics in smart home network,

    M. Fu, P. Wang, S. Liu, X. Chen, and X. Zhou, “Fir-gnn: A graph neural network using flow interaction relationships for intrusion detection of consumer electronics in smart home network,”IEEE Transactions on Consumer Electronics, vol. 71, no. 2, pp. 4892–4902, May 2025

    work page 2025

  32. [32]

    Deep graph infomax,

    P. Veli ˇckovi´c, W. Fedus, W. L. Hamilton, P. Li `o, Y . Bengio, and D. Hjelm, “Deep graph infomax,” inInternational Conference on Learning Representations (ICLR), May 2019

    work page 2019

  33. [33]

    Contrastive multi-view represen- tation learning on graphs,

    K. Hassani and A. H. Khasahmadi, “Contrastive multi-view represen- tation learning on graphs,” inProceedings of the 37th International Conference on Machine Learning (ICML). PMLR, 13-18 Jul 2020, pp. 4116–4126

    work page 2020

  34. [34]

    Large-scale representation learning on graphs via bootstrapping,

    S. Thakoor, C. Tallec, M. G. Azar, M. Azabou, E. L. Dyer, R. Munos, P. Veliˇckovi´c, and M. Valko, “Large-scale representation learning on graphs via bootstrapping,” inInternational Conference on Learning Representations (ICLR), 2022

    work page 2022

  35. [35]

    Handling low homophily in recommender systems with partitioned graph transformer,

    T. T. Nguyen, T. T. Nguyen, M. Weidlich, J. Jo, Q. V . H. Nguyen, H. Yin, and A. W.-C. Liew, “Handling low homophily in recommender systems with partitioned graph transformer,”IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 1, pp. 334–350, Jan 2025

    work page 2025

  36. [36]

    Revisiting adversarial attacks on graph neural networks for graph classification,

    X. Wang, H. Chang, B. Xie, T. Bian, S. Zhou, D. Wang, Z. Zhang, and W. Zhu, “Revisiting adversarial attacks on graph neural networks for graph classification,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 5, pp. 2166–2178, May 2024

    work page 2024

  37. [37]

    Hard negative mixing for contrastive learning,

    Y . Kalantidis, M. B. Sariyildiz, N. Pion, P. Weinzaepfel, and D. Larlus, “Hard negative mixing for contrastive learning,” inAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2020, pp. 21 798–21 809

    work page 2020

  38. [38]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Rep- resentations (ICLR), 2017

    work page 2017

  39. [39]

    Collective classification in network data,

    P. Sen, G. Namata, M. Bilgic, L. Getoor, B. Galligher, and T. Eliassi- Rad, “Collective classification in network data,”AI Magazine, vol. 29, no. 3, p. 93, Sep. 2008

    work page 2008

  40. [40]

    Wiki-CS: A wikipedia-based benchmark for graph neural networks,

    P. Mernyei and C. Cangea, “Wiki-CS: A wikipedia-based benchmark for graph neural networks,” 2022

    work page 2022

  41. [41]

    Pitfalls of graph neural network evaluation,

    O. Shchur, M. Mumme, A. Bojchevski, and S. G ¨unnemann, “Pitfalls of graph neural network evaluation,” 2019

    work page 2019

  42. [42]

    Geom-GCN: Geometric graph convolutional networks,

    H. Pei, B. Wei, K. C.-C. Chang, Y . Lei, and B. Yang, “Geom-GCN: Geometric graph convolutional networks,” inInternational Conference on Learning Representations (ICLR), 2020

    work page 2020

  43. [43]

    Augmentation-free self-supervised learning on graphs,

    N. Lee, J. Lee, and C. Park, “Augmentation-free self-supervised learning on graphs,”Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 36, no. 7, pp. 7372–7380, Jun. 2022

    work page 2022

  44. [44]

    Augmentation-free graph contrastive learning with performance guarantee,

    H. Wang, J. Zhang, Q. Zhu, and W. Huang, “Augmentation-free graph contrastive learning with performance guarantee,” 2022

    work page 2022

  45. [45]

    Graph representation learning with adaptive metric,

    C.-Y . Zhang, H.-C. Cai, C. L. P. Chen, Y .-N. Lin, and W.-P. Fang, “Graph representation learning with adaptive metric,”IEEE Transactions on Network Science and Engineering, vol. 10, no. 4, pp. 2074–2085, July 2023

    work page 2074

  46. [46]

    Learning graph augmentations to learn graph representations,

    K. Hassani and A. H. Khasahmadi, “Learning graph augmentations to learn graph representations,” 2022

    work page 2022

  47. [47]

    COSTA: Covariance-preserving feature augmentation for graph contrastive learn- ing,

    Y . Zhang, H. Zhu, Z. Song, P. Koniusz, and I. King, “COSTA: Covariance-preserving feature augmentation for graph contrastive learn- ing,” inProceedings of the 28th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining (KDD). ACM, 2022, pp. 2524–2534

    work page 2022

  48. [48]

    Unsupervised structure- adaptive graph contrastive learning,

    H. Zhao, X. Yang, C. Deng, and D. Tao, “Unsupervised structure- adaptive graph contrastive learning,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–13, 2023

    work page 2023

  49. [49]

    Graph self- supervised learning with augmentation-aware contrastive learning,

    D. Chen, X. Zhao, W. Wang, Z. Tan, and W. Xiao, “Graph self- supervised learning with augmentation-aware contrastive learning,” in Proceedings of the ACM Web Conference 2023 (WWW). ACM, 2023, pp. 154–164. BIOGRAPHIES Adnan Alireceived his PhD from the School of Computer Science and Technology, University of Science and Technology of China (USTC), Hefei, J...

    work page 2023