Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Model-Driven Graph Contrastive Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MGCL estimates the graphon behind observed graphs, resamples edges from it for contrastive views, and posts the lowest average rank on six node-level and eight graph-level benchmarks.

desk verdict Node-level graphon-informed augmentation is credible; the graph-level SOTA claim rests on an unvalidated clustering assumption that the paper's own synthetic experiment undercuts. read the letter →

arxiv 2506.06212 v1 pith:Q5O5PG36 submitted 2025-06-06 cs.LG

Ali Azizpour , Nicolas Zilberstein , Santiago Segarra This is my paper
classification cs.LG
keywords graphcontrastivelearninggraphongraphon-informedaugmentationself-supervisednodeclassificationfalsenegativesgenerativemodels
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

This paper argues that the weak point of graph contrastive learning is the augmentation step: hand-designed perturbations such as random edge drops ignore the distribution that actually produced the data. MGCL instead estimates a graphon — a bounded symmetric function on the unit square that acts as a probabilistic generative model for graphs — from the observed data and resamples edges according to that estimate to build augmented views. For graph-level tasks it clusters the dataset, estimates one graphon per cluster, and contrasts each graph only against graphs from other clusters, which reduces false negatives. Across six node-level and eight graph-level benchmarks, MGCL posts the lowest average rank among all compared methods (1.67 and 1.75, respectively), which the paper takes as evidence that generative models belong inside the contrastive loop. If that holds, augmentation design becomes a statistical estimation problem rather than a menu of heuristics.

What carries the argument

The load-bearing object is the graphon-informed augmentation transform $T_{\text{GIA}}$. A graphon is a bounded, symmetric, measurable function $W: [0,1]^2 \to [0,1]$ that generates a graph by giving each node an independent latent variable in $[0,1]$ and connecting each pair $(i,j)$ with probability $W(\eta_i,\eta_j)$; MGCL's twist is to invert that generative story — estimate $\widehat{W}$ and the latent positions from data, then treat $\widehat{W}$ as the augmentation distribution by resampling a chosen fraction of edges as Bernoulli draws from $\widehat{W}(\eta_i,\eta_j)$. The companion mechanism for graph-level tasks is model-aware contrasting: each cluster receives its own graphon, two views are drawn per graph from its cluster's graphon, and the loss pulls a graph toward its own first view while pushing away only the second views of graphs in other clusters, so the contrastive signal is defined at the level of generative models rather than individual graphs.

What would settle it

On a synthetic dataset with known ground-truth graphons, such as the paper's own four-graphon construction, measure the purity of the k-means clusters: if same-graphon graphs are not grouped together yet graph classification accuracy still holds, the clustering assumption must be separated from the contrastive loss as the source of the gain. A second check is to replace every estimated graphon with the constant $1/2$ while keeping the cluster structure intact; if accuracy barely moves, the learned edge probabilities are not what carries the result.

Watch

Extended reading notes

Core claim

MGCL's central claim is that the generative process behind graph data can be estimated and then reused as the augmentation mechanism, replacing hand-crafted perturbations. Assuming the observed graphs are samples from an unknown graphon $W$, MGCL fits $W$ along with latent node positions $\eta$ using an implicit neural representation (the SIGL estimator), producing an edge-probability map $f_\varphi(\cdot,\cdot)$. The graphon-informed augmentation $T_{\text{GIA}}$ resamples a random $r\%$ of adjacency entries as Bernoulli draws from that learned map, so each augmented view is a sample from a distribution that respects the data's own generative structure; standard random edge perturbation is the special case in which the graphon is taken to be the constant $1/2$. For graph-level tasks, MGCL partitions the dataset into $K = \log(L)$ clusters via k-means on embeddings from a randomly initialized GNN, estimates a separate graphon per cluster, and optimizes a modified InfoNCE loss whose negative views come only from other clusters. The paper reports average ranks of 1.67 over six node classification datasets and 1.75 over eight graph classification datasets, and a synthetic experiment in which MGCL (86.38% accuracy) nearly closes the gap to a supervised GCN (90.56%) while an equivalent model with random edge perturbations trails at 80.81%.

Load-bearing premise

For graph-level tasks, the method assumes that k-means on embeddings from a randomly initialized GNN groups graphs generated by the same underlying graphon into the same cluster; if that grouping is inaccurate, the per-cluster graphons become noisy and the contrastive loss contrasts unrelated graphs or fails to separate distinct generative structures.

Editorial extensions

If this is right

  • Random edge perturbation becomes a special case of graphon-informed augmentation — the constant-$1/2$ graphon — so the method subsumes the standard heuristic as the least-informed choice.
  • Contrasting graphs cluster-by-cluster instead of instance-by-instance raises the true-negative to false-negative ratio on every dataset tested, with the largest gains (up to 9.2% on COLLAB) appearing on social networks with heterogeneous generative structure.
  • A single estimated graphon improves node classification across citation, co-purchase, and co-authorship networks, reaching the best or second-best accuracy on five of six datasets, including the two largest.
  • The resampling ratio acts as a tunable control: accuracy peaks near $r \in [0.2, 0.4]$, showing that partial structure-aware perturbation beats both no perturbation and full resampling.

Reading between the lines

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

  • The same recipe — estimate a cheap generative prior, then resample from it to build views — could transfer to other self-supervised regimes, such as image or time-series learning, wherever the estimator is faithful enough that the augmentation distribution tracks the real data distribution.
  • If per-cluster graphons are accurate, graph classification accuracy should track cluster purity, so a practitioner could tune the number of clusters by agreement between cluster assignments and known semantic classes rather than by accuracy alone.
  • The paper's own stated ceiling — graphons model edge structure but not node or edge features — suggests that a feature-aware generative model is the natural next step; a testable extension would be whether such a model further reduces false negatives on attributed datasets such as NCI1.
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 MGCL, a model-driven graph contrastive learning framework that uses graphons as generative priors. For node-level tasks, MGCL estimates a graphon from the input graph with SIGL and defines a graphon-informed augmentation (GIA) that resamples a subset of edges from the estimated graphon; the encoder is trained with a DGI objective. For graph-level tasks, MGCL clusters the dataset by k-means on embeddings from a randomly initialized GNN, estimates a per-cluster graphon, generates two GIA views per graph, and trains with a modified InfoNCE loss whose negatives are restricted to graphs in different clusters. The paper reports state-of-the-art average ranks on node classification (A.R. 1.67) and graph classification (A.R. 1.75), node clustering gains, and increased TFR across eight TUDatasets.

Significance. The core idea is timely and interesting: replacing heuristic augmentations with data-driven generative-model-guided augmentations is a principled direction, and the node-level results are consistently strong across six benchmarks. The paper also provides useful ablations on the resampling ratio, the number of clusters, and the number of augmentations, plus a synthetic study of graphon estimation. If the graph-level mechanism is validated, the work would be a solid contribution to self-supervised graph learning. The main caveat is that the graph-level claim of false-negative reduction rests on an unvalidated clustering assumption, and the TFR evidence is partly by construction; these need to be addressed before the claim is fully supported.

major comments (4)
  1. [§3.2, Eq. (6), Appendix D.4] The graph-level mechanism depends on the assumption that k-means over embeddings from a randomly initialized GNN groups graphs generated by the same graphon, but the paper provides no direct validation of this assumption on the real TUDatasets used in Table 4. The only direct evidence, Appendix D.4, actually shows the opposite tendency: with four true graphons and K=7, three true graphon groups are split into two subclusters, so Eq. (6) actively contrasts graphs generated by the same graphon and reintroduces exactly the false negatives the method claims to avoid. Because the negative set in Eq. (6) is defined by c(t), the reported graph-level gains (A.R. 1.75) cannot be attributed to the model-driven mechanism unless the clusters are shown to be semantically meaningful. I recommend adding a cluster-quality analysis on the eight real datasets (e.g., cluster purity/ARI w.r.t. class labels) and an ablation in which the same graphon pipeline is run with random cluster assignments or with an equal-sized random subset of negatives.
  2. [Figure 3, TFR analysis] The TFR improvement is largely by construction: MGCL removes all same-cluster graphs from the negative set, so the negative set is smaller and its class distribution is different from that of the full-batch baseline. Consequently, an increased TFR does not establish that the clusters correspond to generative structure or to the class structure used to define false negatives. The comparison in Figure 3 should be made against a baseline that uses the same negative-set size but selects negatives randomly; otherwise the plot is descriptive of the loss design rather than evidence of semantic clustering. I also note the caption 'Baseline' is used to mean all InfoNCE-based methods, which is imprecise.
  3. [Table 4 and experimental protocol] Table 4 states that all baseline numbers are taken from [50] rather than re-run under a unified protocol. This is a load-bearing limitation for the graph-level state-of-the-art claim, because differences in evaluation code, data splits, and hyperparameters can change rankings in TUDataset benchmarks. The authors should either re-run the baselines under the same protocol or clearly report which implementations and settings were used, and discuss the sensitivity of the reported A.R. to these differences. The absence of released code further limits the reproducibility of both the graphon estimation and the clustering pipeline.
  4. [§3.2, Step 1 and hyperparameters] Several free parameters in the graph-level pipeline are not justified by the experiments: the cluster count K=log(L), the number J=10 of closest graphs used for graphon estimation, and the use of all-ones node features for the initial clustering step. The paper shows some sensitivity to K in Appendix D.2, but J and the feature choice are not varied. Since these choices directly affect the estimated graphons and the negative set in Eq. (6), the authors should either provide a sensitivity analysis or state more explicitly the range of settings under which the reported gains persist.
minor comments (6)
  1. [Abstract/Introduction] The phrase 'graph contrast learning' should be 'graph contrastive learning'.
  2. [§4.2, Table 4] The table caption should be supplemented in the main text with the fact that baseline numbers are borrowed from [50], since the current reader could otherwise assume all methods were run under the same protocol.
  3. [§3.2, Eq. (6)] The notation c(t) is used before it is formally defined; please move the cluster-assignment definition to just before Eq. (5) or include it in the equation caption.
  4. [Appendix C.1] The hyperparameter section does not state the number of random seeds or the initialization scheme for the randomly initialized GNN used in graph-level clustering; please add this for reproducibility.
  5. [Figure 3] The figure caption should clarify what 'Baseline' refers to and how the TFR is averaged over batches and datasets.
  6. [§4.2] The sentence 'MGCL ranks first on five out of eight datasets' is consistent with the table, but the phrase 'outperforming all competing baselines' in the following paragraph should be qualified because MGCL does not rank first on all datasets and is third on RDT-M5K.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MGCL's graphon estimates guide augmentations, but the reported accuracy claims are tested against external baselines and do not reduce to fitted inputs.

full rationale

The claimed derivation chain is not circular. MGCL estimates graphons via SIGL and uses them to define graphon-informed augmentations and model-aware negative sets; the paper's central performance claims are benchmark accuracies against external baselines (Tables 1 and 4), evaluated on held-out labels with linear and SVM protocols. No prediction is defined in terms of a fitted parameter: the TFR analysis in Figure 3 is an evaluation metric on the negative-set design, and although MGCL excludes same-cluster negatives by construction, the TFR ratio is not guaranteed to increase unless the clusters carry class or generative information, so the observed increases are empirical rather than tautological. The synthetic node-level experiment (Table 3) generates labels from a known graphon and latent-variable process, but MGCL never sees those labels during pretraining; it tests the proposed mechanism rather than encoding the outcome. The reliance on SIGL [4] is a self-citation, but SIGL is a separately published graphon estimator with its own assumptions, and Appendix D.4 evaluates graphon recovery independently on synthetic data; it is not used to assert the paper's main result. The graph-level clustering via a randomly initialized GNN is an unverified assumption and a correctness or robustness risk, but it is not a circular step because the downstream accuracies could in principle fail if the clusters were arbitrary.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new theoretical entities. It relies on the existing concept of graphons and a graphon estimator from prior work. Free parameters are standard hyperparameters such as resampling ratio and number of clusters.

free parameters (4)
  • r = 0.2
    Resampling ratio in TGIA, set to 20% in node-level experiments; experiments D.1 show optimal between 0.2 and 0.4.
  • K = log(L)
    Number of graph clusters; set to log(L) with L=dataset size, but D.2 shows performance varies with K, so it is effectively a tunable hyperparameter.
  • J = 10
    Number of graphs closest to cluster center used to estimate per-cluster graphon; chosen to reduce noise.
  • tau (temperature) = not specified (follows GraphCL)
    Temperature in InfoNCE loss (Eq. 6); inherited from GraphCL configuration, not tuned here.
assumptions (3)
  • domain assumption The observed graph is a sample from an underlying graphon (node-level), and the dataset is generated from multiple graphons (graph-level).
    The entire method rests on this generative model assumption; in real-world networks, this may not hold.
  • ad hoc to paper Clustering via k-means on embeddings from a randomly initialized GNN groups graphs generated from the same graphon.
    This is crucial for graph-level MGCL; no theoretical or empirical justification is given beyond intuition (Section 3.2, Step 1).
  • domain assumption The estimated graphon accurately approximates the true graphon, so augmentations drawn from it are semantically faithful.
    Graphon estimation (SIGL) is assumed accurate; errors in estimation would propagate to augmentations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Model-Driven Graph Contrastive Learning." pith.science (2026). https://pith.science/paper/Q5O5PG36

@misc{pith2026250606212,
  author       = {Pith},
  title        = {Pith review of: Model-Driven Graph Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q5O5PG36}},
  note         = {Machine review of arXiv:2506.06212}
}
abstract

We propose $\textbf{MGCL}$, a model-driven graph contrastive learning (GCL) framework that leverages graphons (probabilistic generative models for graphs) to guide contrastive learning by accounting for the data's underlying generative process. GCL has emerged as a powerful self-supervised framework for learning expressive node or graph representations without relying on annotated labels, which are often scarce in real-world data. By contrasting augmented views of graph data, GCL has demonstrated strong performance across various downstream tasks, such as node and graph classification. However, existing methods typically rely on manually designed or heuristic augmentation strategies that are not tailored to the underlying data distribution and operate at the individual graph level, ignoring similarities among graphs generated from the same model. Conversely, in our proposed approach, MGCL first estimates the graphon associated with the observed data and then defines a graphon-informed augmentation process, enabling data-adaptive and principled augmentations. Additionally, for graph-level tasks, MGCL clusters the dataset and estimates a graphon per group, enabling contrastive pairs to reflect shared semantics and structure. Extensive experiments on benchmark datasets demonstrate that MGCL achieves state-of-the-art performance, highlighting the advantages of incorporating generative models into GCL.

Figures

Figures reproduced from arXiv: 2506.06212 by the authors.

Figure 1
Figure 1. Using the underlying graphon to inform the augmentation. In this work, we propose MGCL, a Model-driven Graph Contrastive Learning framework that explicitly incorpo￾rates the underlying generative process into contrastive learning. MGCL assumes that graphs are samples from a shared, but unknown, graphon – a nonparametric prob￾abilistic model for generating graphs [24, 10, 14], which has shown success in various appli… view at source ↗
Figure 2
Figure 2. An overview of MGCL. (a) For node-level tasks, MGCL estimates the underlying graphon to guide augmentation. Two views are sampled from TGIA, with the edge between nodes 3 and 4 highlighted as an example. (b) For a set of graphs, MGCL first clusters the graphs (each white point) and estimates a graphon for each cluster. Two augmentations are then generated for each graph based on its corresponding graphon. Finally, e… view at source ↗
Figure 3
Figure 3. Effect of clustering on TFR across different datasets. Effect of model-aware clustering on false negative reduction. To evaluate the ef￾fect of clustering on the rate of false nega￾tives, we define the True Negative to False Negative Ratio (TFR). To compute this met￾ric, in each data batch, we examine the neg￾ative samples relative to a graph i. Among these, the negative samples that share the same class as graph i … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Cluster-specific estimated graphons in the COLLAB dataset, revealing diverse structures. a fair comparison across methods. We compute the TFR for each graph in the batch and then average it across all graphs in the dataset. As shown in [PITH_FULL_IMAGE:figures/full_fi…
Figure 5
Figure 5. Figure 5: Simulating a graph, its initial features, and node labels using a graphon neural network. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Effect of graphon-based resampling ratio [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Effect of the number of clusters on graph classification performance. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Initial graph embeddings colored by true graphon (left) and MGCL cluster assignments [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Estimated graphons compared to the ground-truth graphons. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

  1. [50]

    You, Y ., Chen, T., Shen, Y ., and Wang, Z. (2021). Graph contrastive learning automated. InIntl. Conf. on Machine Learning (ICML), pages 12121–12132. PMLR

    work page 2021

  2. [1]

    Adhikari, B., Zhang, Y ., Ramakrishnan, N., and Prakash, B. A. (2018). Sub2vec: Feature learning for subgraphs. InAdv. Knowl. Discov. Data Min. Pac.-Asia Conf., pages 170–182. Springer

    work page 2018

  3. [2]

    M., Costa, T

    Airoldi, E. M., Costa, T. B., and Chan, S. H. (2013). Stochastic blockmodel approximation of a graphon: Theory and consistent estimation.Advances in Neural Inf. Process. Syst. (NIPS), 26

    work page 2013

  4. [3]

    J., and Segarra, S

    Azizpour, A., Balaji, A., Treangen, T. J., and Segarra, S. (2024). Graph-based self-supervised learning for repeat detection in metagenomic assembly.Genome Res., 34(9):1468–1476

    work page 2024

  5. [4]

    Azizpour, A., Zilberstein, N., and Segarra, S. (2025). Scalable implicit graphon learning. InInt. Conf. on Artif. Intell. and Stat

    work page 2025

  6. [5]

    L., Fu, Y ., Nute, M

    Balaji, A., Sapoval, N., Seto, C., Elworth, R. L., Fu, Y ., Nute, M. G., Savidge, T., Segarra, S., and Treangen, T. J. (2022). Komb: K-core based de novo characterization of copy number variation in microbiomes.Comput. Struct. Biotechnol. J., 20:3208–3222

    work page 2022

  7. [6]

    Bielak, P., Kajdanowicz, T., and Chawla, N. V . (2022). Graph barlow twins: A self-supervised representation learning framework for graphs.Knowl. Based Syst., 256:109631

    work page 2022

  8. [7]

    and Airoldi, E

    Chan, S. and Airoldi, E. (2014). A consistent histogram estimator for exchangeable graph models. InIntl. Conf. on Machine Learning (ICML), pages 208–216. PMLR

    work page 2014

Show all 56 references
  1. [8]

    Chien, E., Li, M., Aportela, A., Ding, K., Jia, S., Maji, S., Zhao, Z., Duarte, J., Fung, V ., Hao, C., Luo, Y ., Milenkovic, O., Pan, D., Segarra, S., and Li, P. (2024). Opportunities and challenges of graph neural networks in electrical engineering.Nat. Rev. Electr . Eng., 1...

    2024

  2. [9]

    Chowdhury, A., Verma, G., Rao, C., Swami, A., and Segarra, S. (2021). Unfolding wmmse using graph neural networks for efficient power allocation.IEEE Trans. Wireless Commun., 20(9):6004–6017

    2021

  3. [10]

    and Janson, S

    Diaconis, P. and Janson, S. (2007). Graph limits and exchangeable random graphs.arXiv preprint arXiv:0712.2749

    2007 arXiv

  4. [11]

    and Caines, P

    Gao, S. and Caines, P. E. (2019). Graphon control of large-scale networks of linear systems. IEEE Trans. Autom. Control., 65(10):4090–4105

    2019

  5. [12]

    Ghose, A., Zhang, Y ., Hao, J., and Coates, M. (2023). Spectral augmentations for graph contrastive learning. InInt. Conf. on Artif. Intell. and Stat., pages 11213–11266. PMLR

    2023

  6. [13]

    and Leskovec, J

    Grover, A. and Leskovec, J. (2016). node2vec: Scalable feature learning for networks. InProc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pages 855–864

    2016

  7. [14]

    Han, X., Jiang, Z., Liu, N., and Hu, X. (2022). G-mixup: Graph data augmentation for graph classification. InIntl. Conf. on Machine Learning (ICML), pages 8230–8248. PMLR

    2022

  8. [15]

    and Khasahmadi, A

    Hassani, K. and Khasahmadi, A. H. (2020). Contrastive multi-view representation learning on graphs. InIntl. Conf. on Machine Learning (ICML), pages 4116–4126. PMLR

    2020

  9. [16]

    He, D., Shan, L., Zhao, J., Zhang, H., Wang, Z., and Zhang, W. (2024). Exploitation of a latent mechanism in graph contrastive learning: Representation scattering.Advances in Neural Inf. Process. Syst. (NIPS), 37:115351–115376. 10

    2024

  10. [17]

    K., Eberl, T., Panasyuk, A., Netravali, R., Chandra, R., and Kandula, S

    Hsieh, K., Wong, M., Segarra, S., Mani, S. K., Eberl, T., Panasyuk, A., Netravali, R., Chandra, R., and Kandula, S. (2024). NetVigil: Robust and Low-Cost anomaly detection for East-West data center security. InUSENIX Symp. Netw. Syst. Des. Implement. (NSDI), pages 1771–1789

    2024

  11. [18]

    Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V ., and Leskovec, J. (2020). Strategies for pre-training graph neural networks.Intl. Conf. Learn. Repr . (ICLR)

    2020

  12. [19]

    Kingma, D. P. (2015). Adam: A method for stochastic optimization.Intl. Conf. Learn. Repr . (ICLR)

    2015

  13. [20]

    Kipf, T. N. and Welling, M. (2017). Semi-supervised classification with graph convolutional networks.Intl. Conf. Learn. Repr . (ICLR)

    2017

  14. [21]

    Kumar, S., Mallik, A., Khetarpal, A., and Panda, B. S. (2022). Influence maximization in social networks using graph embedding and graph neural network.Inf. Sci., 607:1617–1636

    2022

  15. [22]

    Liu, N., Wang, X., Bo, D., Shi, C., and Pei, J. (2022a). Revisiting graph contrastive learning from the perspective of graph spectrum.Advances in Neural Inf. Process. Syst. (NIPS), 35:2972–2983

    2022

  16. [23]

    Liu, Y ., Jin, M., Pan, S., Zhou, C., Zheng, Y ., Xia, F., and Yu, P. S. (2022b). Graph self- supervised learning: A survey.IEEE Trans. Knowl. Data Eng., 35(6):5879–5900

    2022

  17. [24]

    (2012).Large networks and graph limits, volume 60

    Lovász, L. (2012).Large networks and graph limits, volume 60. American Mathematical Soc

    2012

  18. [25]

    Mémoli, F. (2011). Gromov–Wasserstein distances and the metric approach to object matching. F ound. Comput. Math., 11:417–487

    2011

  19. [26]

    M., Bause, F., Kersting, K., Mutzel, P., and Neumann, M

    Morris, C., Kriege, N. M., Bause, F., Kersting, K., Mutzel, P., and Neumann, M. (2020). Tudataset: A collection of benchmark datasets for learning with graphs. InICML 2020 Workshop on Graph Representation Learning and Beyond (GRL+ 2020)

    2020

  20. [27]

    Narayanan, A., Chandramohan, M., Venkatesan, R., Chen, L., Liu, Y ., and Jaiswal, S. (2017). graph2vec: Learning distributed representations of graphs.arXiv preprint arXiv:1707.05005

    2017 arXiv

  21. [28]

    and Segarra, S

    Navarro, M. and Segarra, S. (2022). Joint network topology inference via a shared graphon model.IEEE Trans. Signal Process., 70:5549–5563

    2022

  22. [29]

    and Segarra, S

    Navarro, M. and Segarra, S. (2023). GraphMAD: Graph mixup for data augmentation using data-driven convex clustering. InIEEE Intl. Conf. Acoust., Speech and Signal Process. (ICASSP), pages 1–5. IEEE

    2023

  23. [30]

    Oord, A. v. d., Li, Y ., and Vinyals, O. (2018). Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748

    2018 arXiv

  24. [31]

    and Ozdaglar, A

    Parise, F. and Ozdaglar, A. (2023). Graphon games: A statistical framework for network games and interventions.Econometrica, 91(1):191–225

    2023

  25. [32]

    Peng, Z., Huang, W., Luo, M., Zheng, Q., Rong, Y ., Xu, T., and Huang, J. (2020). Graph representation learning via graphical mutual information maximization. InProc. The Web Conf., pages 259–270

    2020

  26. [33]

    M., Navarro, M., and Segarra, S

    Roddenberry, T. M., Navarro, M., and Segarra, S. (2021). Network topology inference with graphon spectral penalties. InIEEE Intl. Conf. Acoust., Speech and Signal Process. (ICASSP), pages 5390–5394. IEEE

    2021

  27. [34]

    Rong, Y ., Huang, W., Xu, T., and Huang, J. (2020). DropEdge: Towards deep graph convolu- tional networks on node classification. InIntl. Conf. Learn. Repr . (ICLR)

    2020

  28. [35]

    Ruiz, L., Chamon, L., and Ribeiro, A. (2020). Graphon neural networks and the transferability of graph neural networks.Advances in Neural Inf. Process. Syst. (NIPS), 33:1702–1712

    2020

  29. [36]

    Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T. (2008). Collective classification in network data.AI Mag., 29(3):93–93. 11

    2008

  30. [37]

    Shchur, O., Mumme, M., Bojchevski, A., and Günnemann, S. (2018). Pitfalls of graph neural network evaluation.arXiv preprint arXiv:1811.05868

    2018 arXiv

  31. [38]

    Sitzmann, V ., Martel, J., Bergman, A., Lindell, D., and Wetzstein, G. (2020). Implicit neural representations with periodic activation functions.Advances in Neural Inf. Process. Syst. (NIPS), 33:7462–7473

    2020

  32. [39]

    Sun, F.-Y ., Hoffmann, J., Verma, V ., and Tang, J. (2020). Infograph: Unsupervised and semi- supervised graph-level representation learning via mutual information maximization.Intl. Conf. Learn. Repr . (ICLR)

    2020

  33. [40]

    Suresh, S., Li, P., Hao, C., and Neville, J. (2021). Adversarial graph augmentation to improve graph contrastive learning.Advances in Neural Inf. Process. Syst. (NIPS), 34:15920–15933

    2021

  34. [41]

    G., Munos, R., Veliˇckovi´c, P., and Valko, M

    Thakoor, S., Tallec, C., Azar, M. G., Munos, R., Veliˇckovi´c, P., and Valko, M. (2021). Boot- strapped representation learning on graphs. InICLR workshop on geometrical and topological representation learning

    2021

  35. [42]

    L., Liò, P., Bengio, Y ., and Hjelm, R

    Veliˇckovi´c, P., Fedus, W., Hamilton, W. L., Liò, P., Bengio, Y ., and Hjelm, R. D. (2019). Deep graph infomax.Intl. Conf. Learn. Repr . (ICLR)

    2019

  36. [43]

    Wang, R., Wang, X., Shi, C., and Song, L. (2022). Uncovering the structural fairness in graph contrastive learning.Advances in Neural Inf. Process. Syst. (NIPS), 35:32465–32473

    2022

  37. [44]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Yu, P. S. (2020). A comprehensive survey on graph neural networks.IEEE Trans. Neural Netw. Learn. Syst., 32(1):4–24

    2020

  38. [45]

    Xia, X., Mishne, G., and Wang, Y . (2023). Implicit graphon neural representation.Int. Conf. on Artif. Intell. and Stat., pages 10619–10634

    2023

  39. [46]

    Xu, H., Luo, D., Carin, L., and Zha, H. (2021). Learning graphons via structured gromov- wasserstein barycenters. InProc. AAAI Conf. Artif. Intell., pages 10505–10513

    2021

  40. [47]

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. (2019). How powerful are graph neural networks? Intl. Conf. Learn. Repr . (ICLR)

    2019

  41. [48]

    Xu, K., Li, C., Tian, Y ., Sonobe, T., Kawarabayashi, K.-i., and Jegelka, S. (2018). Representation learning on graphs with jumping knowledge networks. InIntl. Conf. on Machine Learning (ICML), pages 5453–5462. PMLR

    2018

  42. [49]

    Ying, Z., You, J., Morris, C., Ren, X., Hamilton, W., and Leskovec, J. (2018). Hierarchical graph representation learning with differentiable pooling.Advances in Neural Inf. Process. Syst. (NIPS), 31

    2018

  43. [51]

    You, Y ., Chen, T., Sui, Y ., Chen, T., Wang, Z., and Shen, Y . (2020). Graph contrastive learning with augmentations.Advances in Neural Inf. Process. Syst. (NIPS), 33:5812–5823

    2020

  44. [52]

    Zhang, X., Tan, Q., Huang, X., and Li, B. (2024). Graph contrastive learning with personalized augmentation.IEEE Trans. Knowl. Data Eng

    2024

  45. [53]

    Zhao, Z., Verma, G., Rao, C., Swami, A., and Segarra, S. (2022). Link scheduling using graph neural networks.IEEE Trans. Wireless Commun., 22(6):3997–4012

    2022

  46. [54]

    Zhu, Y ., Xu, Y ., Yu, F., Liu, Q., Wu, S., and Wang, L. (2020). Deep Graph Contrastive Representation Learning. InICML Workshop on Graph Representation Learning and Beyond

    2020

  47. [55]

    Zhu, Y ., Xu, Y ., Yu, F., Liu, Q., Wu, S., and Wang, L. (2021). Graph contrastive learning with adaptive augmentation. InProceedings of the web conference 2021, pages 2069–2080

    2021

  48. [56]

    Effect of Graphon-Informed Augmentation (GIA)

    Zhuo, J., Lu, Y ., Ning, H., Fu, K., He, D., Wang, C., Guo, Y ., Wang, Z., Cao, X., Yang, L., et al. (2024). Unified graph augmentations for generalized contrastive learning on graphs.Advances in Neural Inf. Process. Syst. (NIPS), 37:37473–37503. 12 A Algorithm We present the ...

    2024

Pith tools

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