Pith. sign in

REVIEW 2 major objections 5 minor 54 references

Subgraph Gaussian Embedding Contrast for Self-Supervised Graph Representation Learning

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read SubGEC maps graph subgraphs into a structured Gaussian space and contrasts them with optimal transport distances, claiming improved self-supervised graph representation learning.

desk verdict The empirical recipe is sound and worth reporting, but the theoretical proof of Theorem 1 is mathematically invalid as written and should be fixed or retracted before publication. read the letter →

arxiv 2505.23529 v2 pith:CC7M2G5P submitted 2025-05-29 cs.LG cs.AI

classification cs.LGcs.AI
keywords subgraphGaussianembeddinggraphrepresentationlearningself-supervisedcontrastiveoptimaltransportWassersteindistanceGromov-Wassersteinnodeclassification
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

Self-supervised graph representation learning usually creates contrastive views by perturbing the graph structure. This paper claims that where those views live in embedding space matters, not just how they are generated. SubGEC samples induced subgraphs, maps them into a structured Gaussian distribution with a small learnable module, and then contrasts the Gaussian embeddings against the original subgraphs using Wasserstein and Gromov-Wasserstein distances inside an InfoNCE loss. The authors argue this controls the distribution of contrastive pairs, avoids the collapse of embeddings, and is justified by a theorem connecting the combined objective to minimizing an expected KL divergence. Empirically, SubGEC reports the best accuracy on the strongly heterophilic Squirrel, Cornell, and Texas datasets and remains competitive on homophilic benchmarks.

What carries the argument

The load-bearing component is the Subgraph Gaussian Embedding (SGE) module: a GraphSAGE layer followed by two separate graph attention networks that emit per-node means and log-variances of a Gaussian distribution. The embedded subgraph features are sampled via the reparameterization trick and pushed toward a standard normal prior with a closed-form KL divergence. Contrast is then defined by the Wasserstein distance (comparing node feature distributions) and the Gromov-Wasserstein distance (comparing intra-graph distance matrices), both scaled by a temperature parameter and inserted into an InfoNCE-style contrastive loss. The final objective is a weighted sum of the two OT losses plus the KL term, and the paper's Theorem 1 is what links this combined loss to the KL-divergence-minimizing claim.

What would settle it

On a trained SubGEC model, evaluate the per-node KL divergence between the encoder's Gaussian and the standard normal prior across a held-out set; if these values are not close to zero, then the substitution in Equation 13 is unsupported and the theorem's conclusion fails.

Watch

Extended reading notes

Core claim

SubGEC's central claim is that forcing the embeddings of subgraphs to follow a Gaussian prior, and then measuring subgraph similarity with optimal transport rather than with standard inner products, yields contrastive representations that transfer better to downstream node classification. The paper presents Theorem 1, which states that as the number of subgraphs grows, jointly minimizing the OT-based InfoNCE loss and the KL regularizer implicitly minimizes the expected KL divergence between the learned posterior $q_\phi(\tilde{X}\vert X,A)$ and the true posterior $p(\tilde{X}\vert X,A)$, while also increasing the mutual information between the input and the latent embedding. On the experimental side, the paper shows accuracy gains concentrated on heterophilic graphs and presents ablations indicating that the KL term, not dropout or a decoder, is responsible for the improvement.

Load-bearing premise

The proof of Theorem 1 assumes that after minimizing the KL regularizer, the encoder's distribution $q_\phi(\tilde{X}\vert X,A)$ is close enough to the Gaussian prior $p(\tilde{X})$ that one can replace the prior by the encoder inside the mutual-information integral; if that approximation does not hold, the theoretical justification collapses and only the experimental results support the method.

Editorial extensions

If this is right

  • Self-supervised node classification on heterophilic graphs—where connected nodes tend to have different labels—can be done without labels and without specialized supervised architectures.
  • The result suggests that the choice of the embedding distribution is a transferable design principle: any contrastive objective could benefit from a regularized latent space.
  • Optimal transport distances as contrastive similarity measures can be combined with variational regularizers, pointing toward a family of OT-based SSL losses.
  • If the running-time analysis holds, the method can scale to graphs with thousands of nodes with modest per-iteration cost.
  • Ablation evidence indicates that KL regularization and OT distances each contribute; removing either reduces performance, especially on heterophilic datasets.

Reading between the lines

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

  • A natural extension is to test whether Gaussianity per se is the driver by replacing the Gaussian prior with a uniform or Laplace prior while keeping the KL-like regularizer, a comparison the paper does not run.
  • The proof of Theorem 1 relies on the tightness of the InfoNCE bound with the number of samples; for small graphs, the theoretical guarantee weakens, so the method's gains on small datasets like Cornell and Texas may depend on empirical rather than theoretical factors.
  • The same Gaussian-embedding plus OT scheme could be applied to graph-level tasks such as graph classification or molecular property prediction, where subgraphs correspond to functional groups.
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

2 major / 5 minor

Summary. The paper proposes SubGEC, a self-supervised graph representation learning method that maps subgraphs into a Gaussian latent space via a VAE-style reparameterization module with KL regularization, and uses Wasserstein and Gromov-Wasserstein distances as similarity measures in an InfoNCE-style contrastive loss. The method is evaluated on eight node-classification benchmarks, showing competitive or superior accuracy, particularly on heterophilic datasets. The authors also claim a theoretical result (Theorem 1) that minimizing the proposed loss implicitly minimizes the expected KL divergence between the learned posterior and the true posterior.

Significance. If the empirical results are reproducible, SubGEC is a valuable contribution to graph contrastive learning, with strong performance on heterophilic graphs and a clean integration of Gaussian embeddings with OT-based contrastive losses. The paper provides extensive ablations (Tables 3 and 4) and sensitivity analyses (Appendix C) that support many design choices. However, the theoretical justification in Section 4.5 is currently invalid, and the claim of a 'theoretical and empirical proof' of the benefits of Gaussian embeddings is not supported. The empirical component alone, while useful, does not substantiate the theoretical framing.

major comments (2)
  1. [Section 4.5 (Eq. 13)] The proof of Theorem 1 is mathematically invalid. In Eq. (13), the step from ∫∫ p(x|z)p(z) log[p(z|x)/p(z)] dx dz to ∫ p(x|z)[∫ q_phi(z|x) log(p(z|x)/q_phi(z|x)) dz] dx replaces the prior p(z) with the variational posterior q_phi(z|x) inside the integral over x. This substitution is unjustified: q_phi depends on x, so it cannot be factored out of the x-integral, and minimizing KL(q_phi(z|x)||p(z)) only yields q_phi ≈ p(z) on average, not pointwise for every x. Moreover, the inner integral equals -KL(q_phi(z|x)||p(z|x)), so the claimed equality would imply I(x,z) = -E[KL] ≤ 0, contradicting the non-negativity of mutual information. The correct identity is I(x,z) = E_{p(x)}[KL(p(z|x)||p(z))], which does not involve q_phi in the stated manner. Consequently, the theoretical justification for the claim that SubGEC minimizes E[KL(q_phi(X̃|X,A)||p(X̃|X,A))] collapses. The authors should either provide a correct proof or explicitly retract the theoretical statements in Section 4.5 and the related claims in the Introduction and Conclusion.
  2. [Section 4.5 (after Eq. 13)] The proof also fails to connect the generic InfoNCE bound of Proposition 1 to the specific losses in Eq. (9). The losses LW and LGW employ distances W and GW with negative pairs drawn from {X_j} ∪ {X̃_j}, and the similarity is an exponentiated negative distance rather than a standard inner product. The paper does not show that these losses satisfy the conditions of Proposition 1 (or a suitable variant), so even if the KL substitution were valid, the chain 'minimizing LW ⇒ maximizing I(X,X̃)' is not established. Please provide a rigorous derivation or remove the claim.
minor comments (5)
  1. [Table 3] The symbols '%' and '✓' are not defined in the caption. Please define them explicitly so that the rows are interpretable.
  2. [Implementation details (Section 5)] The text says the code 'are available1 and will be made public after acceptance', which is ambiguous. Please state clearly whether the code is currently available at the given URL.
  3. [Equations (2), (3), and (9)] The temperature τ is used both in the distance kernel and as a scaling factor for W and GW. Please specify whether the same τ is intended and justify this double scaling.
  4. [Table 4] The statement that excluding OT distances leads to suboptimal performance on heterophilic datasets is not universally true; for example, the L1-only model achieves 90.33 on Cornell, above the GW-only model. Please refine the claim.
  5. [Section 1 / Figure 1] The characterization of SubGEC's embedding space as 'dense, uniform, and linearly separable' is based on qualitative t-SNE plots. Consider adding quantitative metrics (e.g., silhouette score, number of collapsed dimensions) or tempering the wording.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical results are genuine held-out evaluations, and the flawed theoretical proof is a correctness defect rather than a circular reduction.

full rationale

The paper's central empirical claim is that SubGEC outperforms or matches state-of-the-art methods on eight benchmarks. These results are reported on standard test splits after hyperparameters are tuned on validation sets, which is a normal and non-circular procedure: no fitted parameter is renamed as a prediction, and the reported accuracies are not used to define the loss. The method's loss in Eq. (10) combines OT-based InfoNCE terms with a KL regularization term; each component is specified independently of the benchmark outcomes. The theoretical result, Theorem 1, invokes the standard InfoNCE mutual-information bound from Oord et al. [34] and then attempts to relate KL regularization to posterior alignment. The proof in Eq. (13) is mathematically invalid: it replaces the prior p(z) with the variational posterior q_phi(z|x) inside the mutual-information integral without justification, which would imply a negative mutual information. However, this is a correctness defect, not circularity: the theorem's conclusion is not assumed among its premises, and the loss is not defined in terms of the target expected KL divergence. The paper's self-citations ([12] and [38]) appear only as background or hyperparameter-tuning references and are not load-bearing. No uniqueness result from the authors is invoked to rule out alternatives, and no ansatz is imported through self-citation. Therefore, while the theoretical justification deserves scrutiny, the derivation chain does not reduce to its own inputs, and the empirical evaluation stands as independent support for the method's practical claims.

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

The method introduces no new physical or abstract entities. The free parameters are standard hyperparameters tuned on validation. The main assumptions are the Gaussian posterior ansatz and the validity of the InfoNCE bound, plus an ad hoc replacement step in the proof of Theorem 1.

free parameters (4)
  • beta (KL regularization weight) = tuned via random search on validation (e.g., ~1e-3 for Cora)
    Controls the strength of the Gaussian prior penalty. The sensitivity analysis in Appendix C shows performance varies with beta.
  • alpha (balance between Wasserstein and Gromov-Wasserstein losses) = tuned via random search on validation
    Balances feature-distribution and structural fidelity terms in the loss.
  • tau (temperature in OT cost) = tuned via random search on validation
    Used in the exponential cosine distance kernel and in scaling OT distances in the InfoNCE loss.
  • subgraph size k (BFS sampling radius) = sensitivity analyzed on Cora; values 5, 15, 25, 35; default not specified
    Determines the size of the induced BFS subgraphs. Appendix C shows robustness across 5-25 but decline at 35.
assumptions (4)
  • standard math InfoNCE loss is a valid lower bound on mutual information, as proven in Oord et al. [34].
    The theoretical analysis in Section 4.5 relies directly on Proposition 1 from [34].
  • domain assumption The posterior q(\tilde X | X, A) is a product of diagonal Gaussians N(mu_i, diag(sigma_i^2)).
    This Gaussian assumption is the core of the SGE module and is a modeling choice, not derived.
  • ad hoc to paper After KL minimization, q(z|x) is close enough to the prior p(z) that p(z) can be replaced by q(z|x) in the mutual information integral.
    Used in Theorem 1's proof (Section 4.5, Equation 13). This step is not justified and is the weakest link in the theoretical claim.
  • domain assumption The optimal transport cost d(x_m^i, x_n^j) = exp(-cosine_sim(x_m^i, x_n^j) / tau) is a valid dissimilarity measure for subgraph node features.
    The paper defines this kernel in Section 3.3 without proving it is a proper metric; it is used as the ground cost in equations (2) and (3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Subgraph Gaussian Embedding Contrast for Self-Supervised Graph Representation Learning." pith.science (2026). https://pith.science/paper/CC7M2G5P

@misc{pith2026250523529,
  author       = {Pith},
  title        = {Pith review of: Subgraph Gaussian Embedding Contrast for Self-Supervised Graph Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CC7M2G5P}},
  note         = {Machine review of arXiv:2505.23529}
}
read the original abstract

Graph Representation Learning (GRL) is a fundamental task in machine learning, aiming to encode high-dimensional graph-structured data into low-dimensional vectors. Self-Supervised Learning (SSL) methods are widely used in GRL because they can avoid expensive human annotation. In this work, we propose a novel Subgraph Gaussian Embedding Contrast (SubGEC) method. Our approach introduces a subgraph Gaussian embedding module, which adaptively maps subgraphs to a structured Gaussian space, ensuring the preservation of input subgraph characteristics while generating subgraphs with a controlled distribution. We then employ optimal transport distances, more precisely the Wasserstein and Gromov-Wasserstein distances, to effectively measure the similarity between subgraphs, enhancing the robustness of the contrastive learning process. Extensive experiments across multiple benchmarks demonstrate that \method~outperforms or presents competitive performance against state-of-the-art approaches. Our findings provide insights into the design of SSL methods for GRL, emphasizing the importance of the distribution of the generated contrastive pairs.

Figures

Figures reproduced from arXiv: 2505.23529 by the authors.

Figure 1
Figure 1. t-stochastic neighbor embedding (t-SNE) visualizations of previous graph [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SubGEC method. Our model employs a graph encoder [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Average time to compute loss per iteration as a function of the number of [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The plot displays the mean test accuracy (solid blue line) along with a [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 49 canonical work pages

  1. [1]

    1–56 (2024)

    Arya, S., Auddy, A., Clark, R.A., Lim, S., Memoli, F., Packer, D.: The Gromov– Wassersteindistancebetweenspheres.FoundationsofComputationalMathematics pp. 1–56 (2024)

  2. [2]

    In: International Conference on Machine Learning (2022)

    Brogat-Motte, L., Flamary, R., Brouard, C., Rousu, J., d’Alché Buc, F.: Learning to predict graphs with fused Gromov-Wasserstein barycenters. In: International Conference on Machine Learning (2022)

  3. [3]

    Catalogue of Artificial Intelligence Tools pp

    Bundy, A., Wallen, L.: Breadth-first search. Catalogue of Artificial Intelligence Tools pp. 13–13 (1984)

  4. [4]

    In: International Conference on Learning Representations (2024)

    Chen, J., Lei, R., Wei, Z.: PolyGCL: Graph contrastive learning via learnable spec- tral polynomial filters. In: International Conference on Learning Representations (2024)

  5. [5]

    In: International Conference on Machine Learning (2020)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International Conference on Machine Learning (2020)

  6. [6]

    IEEE Access7, 6269–6278 (2018)

    Chen, Y., Georgiou, T.T., Tannenbaum, A.: Optimal transport for Gaussian mix- ture models. IEEE Access7, 6269–6278 (2018)

  7. [7]

    In: International Conference on Learning Representa- tions (2021)

    Chien, E., Peng, J., Li, P., Milenkovic, O.: Adaptive universal generalized PageR- ank graph neural network. In: International Conference on Learning Representa- tions (2021)

  8. [8]

    AAAI Conference on Artificial Intelligence (1998) Subgraph Gaussian Embedding Contrast for Self-Supervised GRL 15

    Craven, M., DiPasquo, D., Freitag, D., McCallum, A., Mitchell, T., Nigam, K., Slattery, S.: Learning to extract symbolic knowledge from the World Wide Web. AAAI Conference on Artificial Intelligence (1998) Subgraph Gaussian Embedding Contrast for Self-Supervised GRL 15

Show all 54 references
  1. [9]

    In: Advances in Neural Information Processing Systems (2019)

    Gasteiger, J., Weiß enberger, S., Günnemann, S.: Diffusion improves graph learn- ing. In: Advances in Neural Information Processing Systems (2019)

  2. [10]

    In: Advances in Neural Information Processing Systems (2016)

    Genevay, A., Cuturi, M., Peyré, G., Bach, F.: Stochastic optimization for large- scale optimal transport. In: Advances in Neural Information Processing Systems (2016)

  3. [11]

    In: ACM Conference on Digital Libraries (1998)

    Giles,C.L.,Bollacker,K.D.,Lawrence,S.:CiteSeer:Anautomaticcitationindexing system. In: ACM Conference on Digital Libraries (1998)

  4. [12]

    In: ACM International Conference on Information and Knowledge Management (2023)

    Giraldo, J.H., Skianis, K., Bouwmans, T., Malliaros, F.D.: On the trade-off be- tween over-smoothing and over-squashing in deep graph neural networks. In: ACM International Conference on Information and Knowledge Management (2023)

  5. [13]

    In: International Conference on Artificial Intelligence and Statistics (2020)

    Goldfeld, Z., Greenewald, K.: Gaussian-smoothed optimal transport: Metric struc- ture and statistical efficiency. In: International Conference on Artificial Intelligence and Statistics (2020)

  6. [14]

    In: Advances in Neural Information Processing Systems (2020)

    Grill, J.B., et al.: Bootstrap your own latent: A new approach to self-supervised learning. In: Advances in Neural Information Processing Systems (2020)

  7. [15]

    In: International Conference on Artificial Intelligence and Statistics (2010)

    Gutmann, M., Hyvärinen, A.: Noise-contrastive estimation: A new estimation prin- ciple for unnormalized statistical models. In: International Conference on Artificial Intelligence and Statistics (2010)

  8. [16]

    In: Advances in Neural Information Processing Systems (2017)

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. In: Advances in Neural Information Processing Systems (2017)

  9. [17]

    In: European Conference on Computer Vision (2022)

    Han, Y., Hui, L., Jiang, H., Qian, J., Xie, J.: Generative subgraph contrast for self- supervised graph representation learning. In: European Conference on Computer Vision (2022)

  10. [18]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018)

    He, A., Luo, C., Tian, X., Zeng, W.: A twofold siamese network for real-time object tracking. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2018)

  11. [19]

    In: Advances in Neural Information Processing Sys- tems (2022)

    He, M., Wei, Z., Wen, J.R.: Convolutional neural networks on graphs with Cheby- shev approximation, revisited. In: Advances in Neural Information Processing Sys- tems (2022)

  12. [20]

    arXiv preprint arXiv:1703.07737 (2017)

    Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re- identification. arXiv preprint arXiv:1703.07737 (2017)

  13. [21]

    In: ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining (2022)

    Hou,Z.,Liu,X.,Cen,Y.,Dong,Y.,Yang,H.,Wang,C.,Tang,J.:GraphMAE:Self- supervised masked graph autoencoders. In: ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining (2022)

  14. [22]

    Technologies9(1), 2 (2020)

    Jaiswal, A., Babu, A.R., Zadeh, M.Z., Banerjee, D., Makedon, F.: A survey on contrastive self-supervised learning. Technologies9(1), 2 (2020)

  15. [23]

    In: IEEE International Conference on Data Mining (2020)

    Jiao, Y., Xiong, Y., Zhang, J., Zhang, Y., Zhang, T., Zhu, Y.: Sub-graph contrast for scalable self-supervised graph representation learning. In: IEEE International Conference on Data Mining (2020)

  16. [24]

    In: International Conference on Learning Representations (2022)

    Jing, L., Vincent, P., LeCun, Y., Tian, Y.: Understanding dimensional collapse in contrastive self-supervised learning. In: International Conference on Learning Representations (2022)

  17. [25]

    Neural Networks (2024)

    Ju, W., Fang, Z., Gu, Y., Liu, Z., Long, Q., Qiao, Z., Qin, Y., Shen, J., Sun, F., Xiao, Z., et al.: A comprehensive survey on deep graph representation learning. Neural Networks (2024)

  18. [26]

    arXiv preprint arXiv:2405.11868 (2024)

    Ju, W., Wang, Y., Qin, Y., Mao, Z., Xiao, Z., Luo, J., Yang, J., Gu, Y., Wang, D., Long, Q., Yi, S., Luo, X., Zhang, M.: Towards graph contrastive learning: A survey and beyond. arXiv preprint arXiv:2405.11868 (2024)

  19. [27]

    Xie et al

    Kingma,D.P.,Welling,M.:Auto-encodingvariationalbayes.In:InternationalCon- ference on Learning Representations (2014) 16 S. Xie et al

  20. [28]

    In: Advances in Neural Information Processing Systems - Workshop (2016)

    Kipf, T.N., Welling, M.: Variational graph auto-encoders. In: Advances in Neural Information Processing Systems - Workshop (2016)

  21. [29]

    In: International Conference on Learning Representations (2017)

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: International Conference on Learning Representations (2017)

  22. [30]

    IEEE Signal Processing Magazine 34(4), 43–59 (2017)

    Kolouri, S., Park, S.R., Thorpe, M., Slepcev, D., Rohde, G.K.: Optimal mass trans- port: Signal processing and machine-learning applications. IEEE Signal Processing Magazine 34(4), 43–59 (2017)

  23. [31]

    IEEE Transactions on Intelligent Transportation Systems 23(3), 1755–1766 (2020)

    Liu, J., Ong, G.P., Chen, X.: GraphSAGE-based traffic speed forecasting for seg- ment network with sparse data. IEEE Transactions on Intelligent Transportation Systems 23(3), 1755–1766 (2020)

  24. [32]

    Liu, Y., Zheng, Y., Zhang, D., Lee, V.C., Pan, S.: Beyond smoothing: Unsupervised graphrepresentationlearningwithedgeheterophilydiscriminating.In:Proceedings of the AAAI Conference on Artificial Intelligence (2023)

  25. [33]

    In: European Conference on Machine Learning and Knowledge Discovery in Databases (2024)

    Liu, Y., Zhang, H., He, T., Zheng, T., Zhao, J.: Bootstrap latents of nodes and neighbors for graph self-supervised learning. In: European Conference on Machine Learning and Knowledge Discovery in Databases (2024)

  26. [34]

    arXiv preprint arXiv:1807.03748 (2018)

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

  27. [35]

    In: International Conference on Learning Representations (2020)

    Pei, H., Wei, B., Chang, K.C.C., Lei, Y., Yang, B.: Geom-GCN: Geometric graph convolutional networks. In: International Conference on Learning Representations (2020)

  28. [36]

    Journal of Complex Networks9(2) (2021)

    Rozemberczki, B., Allen, C., Sarkar, R.: Multi-scale attributed node embedding. Journal of Complex Networks9(2) (2021)

  29. [37]

    Proba- bility Theory and Related Fields (1985)

    Rüschendorf, L.: The Wasserstein distance and approximation theorems. Proba- bility Theory and Related Fields (1985)

  30. [38]

    Transactions on Machine Learning Research (2025)

    SANGARE, A.S., Dunou, N., Giraldo, J.H., Malliaros, F.D.: A fused Gromov- Wasserstein approach to subgraph contrastive learning. Transactions on Machine Learning Research (2025)

  31. [39]

    AI Magazine29(3), 93–93 (2008)

    Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., Eliassi-Rad, T.: Collec- tive classification in network data. AI Magazine29(3), 93–93 (2008)

  32. [40]

    In: Advances in Neural Information Processing Systems - Workshops (2018)

    Shchur, O., Mumme, M., Bojchevski, A., Günnemann, S.: Pitfalls of graph neu- ral network evaluation. In: Advances in Neural Information Processing Systems - Workshops (2018)

  33. [41]

    In: International Conference on Learning Representations (2021)

    Thakoor, S., Tallec, C., Azar, M.G., Azabou, M., Dyer, E.L., Munos, R., Veličković, P., Valko, M.: Large-scale representation learning on graphs via bootstrapping. In: International Conference on Learning Representations (2021)

  34. [42]

    In: International Conference on Machine Learning

    Titouan, V., Courty, N., Tavenard, R., Flamary, R.: Optimal transport for struc- tured data with application on graphs. In: International Conference on Machine Learning. pp. 6275–6284 (2019)

  35. [43]

    In: International Conference on Learning Representations (2022)

    Topping,J.,Giovanni,F.D.,Chamberlain,B.P.,Dong,X.,Bronstein,M.M.:Under- standing over-squashing and bottlenecks on graphs via curvature. In: International Conference on Learning Representations (2022)

  36. [44]

    IEEE Transactions on Information Theory60(7), 3797–3820 (2014)

    Van Erven, T., Harremos, P.: Rényi divergence and Kullback-Leibler divergence. IEEE Transactions on Information Theory60(7), 3797–3820 (2014)

  37. [45]

    Algorithms13(9), 212 (2020)

    Vayer, T., Chapel, L., Flamary, R., Tavenard, R., Courty, N.: Fused Gromov- Wasserstein distance for structured objects. Algorithms13(9), 212 (2020)

  38. [46]

    In: International Conference on Learning Representations (2019)

    Veličković, P., Fedus, W., Hamilton, W.L., Liò, P., Bengio, Y., Hjelm, R.D.: Deep graph infomax. In: International Conference on Learning Representations (2019)

  39. [47]

    In: International Conference on Learning Representations (2018) Subgraph Gaussian Embedding Contrast for Self-Supervised GRL 17

    Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y.: Graph attention networks. In: International Conference on Learning Representations (2018) Subgraph Gaussian Embedding Contrast for Self-Supervised GRL 17

  40. [48]

    Villani, C.: Topics in optimal transportation, vol. 58. American Mathematical Soc. (2021)

  41. [49]

    IEEE Transactions on Neural Networks and Learning Systems 32(1), 4–24 (2020)

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Philip, S.Y.: A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems 32(1), 4–24 (2020)

  42. [50]

    In: ACM International Conference on Information and Knowledge Man- agement (2023)

    Yuan, M., Chen, M., Li, X.: MUSE: Multi-view contrastive learning for heterophilic graphs. In: ACM International Conference on Information and Knowledge Man- agement (2023)

  43. [51]

    In: Advances in Neural Information Processing Systems - Workshops (2023)

    Zhu, J., Xu, K., Tannenbaum, A.: Optimal transport for vector Gaussian mix- ture models. In: Advances in Neural Information Processing Systems - Workshops (2023)

  44. [52]

    In: International Conference on Machine Learning - Workshops (2020)

    Zhu, Y., Xu, Y., Yu, F., Liu, Q., Wu, S., Wang, L.: Deep graph contrastive repre- sentation learning. In: International Conference on Machine Learning - Workshops (2020)

  45. [53]

    In: Proceedings of the Web Conference (2021)

    Zhu, Y., Xu, Y., Yu, F., Liu, Q., Wu, S., Wang, L.: Graph contrastive learning with adaptive augmentation. In: Proceedings of the Web Conference (2021)

  46. [54]

    Zhuo, J., Lu, Y., Ning, H., Fu, K., Niu, B., He, D., Wang, C., Guo, Y., Wang, Z., Cao, X., et al.: Unified graph augmentations for generalized contrastive learning on graphs. In: Advances in Neural Information Processing Systems (2024) A Graph Convolutional Network The graph e...

Pith tools

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