acceptodds
Under review as a conference paper at ICLR 2027

Optimal Transport-Guided Geometric Contrastive Learning for Attributed Graph Clustering

Abstract

Deep attributed graph clustering learns discriminative node representations by fusing node attributes and topology without label supervision, but two issues persist. Propagation operators — fixed spectral convolutions or learned attention scores — aggregate neighbor features by topology alone or by a metrically ungrounded scorer, injecting noise whenever a neighbor is uninformative for a node's cluster identity. The contrastive objectives typically used to sharpen these representations rely on stochastic graph augmentation and large negative pools, both heuristic, costly, and prone to distorting graph semantics through arbitrary perturbation. We introduce OT-GeoCL, which resolves both issues with a single entropic optimal transport coupling. Node attributes are mapped onto the unit hypersphere, and a sparse Sinkhorn coupling, computed from the induced geometric cost and restricted to the graph's edges, serves simultaneously as a geometry-adaptive propagation operator that down-weights uninformative neighbors by construction, and as the soft positive-pair weighting of a contrastive loss requiring neither augmentation nor heuristic negative sampling. We prove the resulting objective admits an exact, non-asymptotic alignment-uniformity decomposition, belongs to the InfoNCE family with explicit limiting cases, and resists representation collapse via spherical potential theory — guarantees unavailable to heavier, heuristically assembled pipelines. Across 12 widely used attributed graph clustering benchmarks spanning citation, co-purchase, and co-authorship networks, OT-GeoCL outperforms fourteen representative generative, contrastive, and multi-task self-supervised baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Entropic Optimal Transport, Contrastive learning, Attributed Graph, Clustering

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_c3KtOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

OT-GeoCL offers a tidy combination of sparse entropic optimal transport and hyperspherical contrastive learning that yields competitive (and on heterophilous data, superior) clustering results under a simple linear encoder. However, the technical novelty is modest, the theoretical claims are largely rearrangements of known facts, the experimental comparison is incomplete and potentially biased by uneven hyperparameter effort, and the overall contribution does not clear the bar for a top-tier venue.

Critical strengths
  1. The unification of propagation and contrastive weighting through a single sparse, doubly-stochastic coupling is clean and conceptually elegant; the exact finite-NN alignment-uniformity decomposition is a neat technical observation that follows directly from that construction.

  2. Empirical performance is competitive on high-homophily graphs and markedly stronger on the low-homophily subset of DGCBench, providing concrete evidence that geometry-aware reweighting of existing edges can mitigate heterophily more effectively than many existing clustering pipelines.

Critical weaknesses
  1. Overstated novelty and incremental technical contribution. The central idea—computing a sparse entropic Sinkhorn coupling restricted to the observed edges and reusing it for both geometry-adaptive message passing and positive-pair weighting—is a modest packaging of existing ingredients. Sinkhorn-normalized attention (SinkGAT), OT-based graph operators, and doubly-stochastic affinities for clustering are already established. Replacing a learned attention scorer with a fixed chordal cost on the hypersphere and feeding the resulting plan into an InfoNCE-style loss does not constitute a fundamental advance. The “single mechanism solves both noisy propagation and augmentation-free contrastive learning” framing is largely rhetorical; the same dual use could be achieved with any doubly-stochastic affinity.

  2. Theoretical claims are largely re-derivations or minor consequences rather than new insight.Theorem 1’s exact finite-NN alignment-uniformity decomposition follows immediately from the double stochasticity of P\mathbf{P} and is a straightforward algebraic rearrangement; the asymptotic version is already classic (Wang & Isola). Proposition 1 merely recovers well-known special cases of InfoNCE / NT-Xent / SupCon under particular choices of the support or temperature. Corollary 1 invokes standard spherical potential theory to argue anti-collapse; this is not specific to the proposed construction. The paper repeatedly advertises these results as “theoretical guarantees not available to heavier, heuristically-assembled pipelines,” which is an overclaim.

  3. Experimental protocol is incomplete and potentially biased.All baseline numbers are taken verbatim from Wu et al. (2025) without re-implementation under identical training budgets, optimizer settings, or embedding dimensions for OT-GeoCL. Hyperparameters (τ\tau, learning rate, weight decay, initial σ\sigma, training length) are tuned per dataset for the proposed method, while baselines use their original defaults. The largest gains appear on the three low-homophily graphs, yet the paper does not compare against any method designed for heterophily. The twelfth DGCBench dataset (ogbn-arXiv) is omitted. The O(N2)O(N^2) full-softmax term is acknowledged but left unaddressed; the method therefore does not scale beyond the modest sizes already present in the benchmark.

  4. Methodological fragility and limited architectural depth.The encoder is a single linear projection followed by ℓ2\ell_2-normalization. The transport support is hard-constrained to the observed (or kk-NN) edges, so under extreme heterophily the method still depends on a potentially noisy topological prior; geometry only reweights within that prior. There is no multi-layer stacking, no residual connections, and no learned cost network—choices that make the model unusually simple but also limit its expressive power relative to modern GNN clustering pipelines. The exploratory heterophilous ablation is insufficiently controlled to isolate the contribution of geometry versus the mere use of a doubly-stochastic operator.

  5. Insufficient significance and overclaiming relative to the venue. The work sits at the intersection of three mature lines (graph clustering, contrastive learning on graphs, and Sinkhorn operators). While the empirical results on low-homophily data are noticeable, the paper does not deliver a new algorithmic paradigm, a substantially weaker set of assumptions, or a scalable solution that would change practice. The repeated emphasis on “exact, non-asymptotic” guarantees and “one mechanism” does not compensate for the incremental nature of the contribution.

Questions for authors

please refer to the weaknesses

Reviewer_vBfnOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OT-GeoCL for attributed graph clustering. It maps node features to the unit sphere and uses a sparse Sinkhorn coupling for propagation and positive-pair weighting. The authors claim an exact alignment-uniformity decomposition, connections to existing losses, and an anti-collapse guarantee. Experiments cover eleven DGCBench datasets.

Critical strengths

This paper presents a simple graph clustering method with large reported gains on low-homophily graphs.

  • Sharing one transport coupling between propagation and positive-pair weighting is an interesting design with few learned components.

  • On Flickr, Table 3 reports 74.09% ACC, compared with 31.51% for the best listed baseline.

Critical weaknesses

Theoretical claims

  • Theorem 1 requires exact row and column sums of 1, but the implementation uses five Sinkhorn iterations. The final column update can change the row sums. Appendix A checks rows before this step. Please report the final marginal errors and their effect on the decomposition.

  • Proposition 1(a) does not recover standard NT-Xent exactly. Eq. (9) includes self-similarity in the denominator, while SimCLR, Eq. (1) excludes it. The self-loop support also needs to change to give one positive per anchor. Please correct the equivalence claim.

  • Corollary 1 applies Cohn and Kumar's pairwise-energy result to a log-sum objective without justifying the extension. The full loss also needs analysis. In the complete-graph, high-temperature limit, P_ij=1/N and collapsed embeddings minimize the loss. The anti-collapse claim needs explicit conditions.

Motivation and experiments

  • DGCBench is from 2025, but its listed baselines date from 2016-2024. Section 6.4 states that none targets heterophily, where the largest gains occur. Please compare with DGAC (WWW 2025), which evaluates both homophilic and heterophilic graphs, using the same data and comparable tuning and selection.

  • Transport only reweights existing edges and self-loops. Are the gains mainly from suppressing propagation? Please report self-loop mass and compare feature-only clustering, the learned encoder with P=I, and row-normalization of the same Gibbs kernel.

  • The paper repeatedly refers to Appendix C for hyperparameters, ablations, and Algorithm 2, but the supplied 14-page manuscript ends with Appendix B. This missing material prevents checking the experimental setup and ablation claims.

  • OT-GeoCL is tuned per dataset, while baseline scores are taken from DGCBench. Please give the search ranges and selection criterion, including whether evaluation labels are used. Report standard deviations for the ten runs.

Questions for authors
  • The listed baselines stop at 2024. Do the gains remain against DGAC (WWW 2025), especially on low-homophily graphs, with the same data and comparable tuning and selection?

  • How does OT-GeoCL compare with feature-only clustering, P=I, and row-normalized-kernel controls? Please provide Appendix C, the selection rule, and standard deviations. Were evaluation labels used for model selection?

  • How are the exact decomposition and NT-Xent equivalence justified for the implemented method? What conditions exclude collapse in the uniform-coupling limiting case?

Reviewer_3YYCOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Aiming at the problem of how to share geometric structure between propagation and contrastive positive samples in attributed graph clustering, this paper proposes OT-GeoCL: using sparse entropy optimal transport coupled with both propagation operator and contrastive positive sample weights on edge supports, and gives the exact alignment-uniform decomposition of this loss. The authors conducted detailed experiments, including ablation experiments, parameter sensitivity analysis, visualization analysis, etc.

Critical strengths

The research motivation is clear, the experimental design is reasonable, the writing is clear and easy to understand. The model shows a large improvement in effect on heterogeneous graphs.

Critical weaknesses

The reference Wu et al. (2025), which is cited multiple times throughout the manuscript, is suspected to be a hallucinated citation generated by a large language model. The authors of the paper published at NeurIPS 2025 do not match those of Wu et al. listed in the references of this paper.

The correct reference should be:

Benyu Wu, Yue Liu, Qiaoyu Tan, Xinwang Liu, Wei Du, Jun Wang, Guoxian Yu. DGCBench: A deep graph clustering benchmark. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS), 2025.

Questions for authors
  1. The ablation experiments should be included in the main text, rather than in the appendix.
  2. It is recommended to supplement the latest methods as baselines for comparison.
  3. Why not validate on the large-scale dataset ogbn-arXiv? Is it because the method does not have scalability for large-scale data?
  4. The reference Wu et al. (2025) mentioned throughout the manuscript is suspected to be a hallucinated citation generated by a large language model. Please verify the authors of the paper published at NeurIPS 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.