acceptodds
Under review as a conference paper at ICLR 2027

Optimal Transport Guided Dual-Distribution Constraints for Multi-view Graph Clustering

Abstract

Multi-view graph clustering aims to group graph nodes by jointly exploiting multi-view node attributes and topological connections. However, existing methods lack reliable global supervision for cross-view representation alignment. Under severe inter-view noise and heterogeneity, representations of each view converge to distinct local optima and yield weak cross-view consensus. Furthermore, most works only regularize node features while neglecting graph topological neighborhood distributions, decoupling graph smoothing from clustering objectives. To address these issues, we propose a novel Optimal Transport Guided Dual-Distribution Constraints for Multi-view Graph Clustering (OT-DDC). It employs an optimal transport-based pseudo-label generation module to construct global cross-view self supervision and alleviate weak cross-view consensus. Meanwhile, the dual-distribution constraint clustering module jointly optimizes cluster centroids in the feature space and topology-aware neighborhood distribution centers, reducing misalignment between semantic representations and graph topological structures. Extensive experiments demonstrate the effectiveness of our algorithm.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view Graph Clustering, Optimal Transport, Dual-Distribution Constraints

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_m4yAOpenReview ↗
Rating
1
Confidence
4
Informativeness
3(+2)
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes OT-DDC, a multi-view attributed graph clustering framework. The method first applies graph Laplacian smoothing to two feature views and uses MLP encoders to obtain view-specific node representations. It then combines topology-guided contrastive learning with cross-view node-level contrastive learning. Its main module employs entropy-regularized optimal transport (OT) to generate soft cross-view pseudo-labels using both node--prototype feature similarity and degree-based topological costs. Finally, the model jointly constrains feature-space cluster assignments and neighborhood-distribution assignments, and applies kk-means to the fused representation. The method is evaluated on AMAP, WIKI, Citeseer, and Cora using ACC, NMI, and ARI. The validation includes comparisons with graph clustering baselines, component ablations, an OT ablation, hyperparameter sensitivity analysis, and t-SNE visualizations.

Critical strengths
  1. The method is intuitive.
  2. The overall presentation is reasonably clear, and the framework figure helps explain the intended pipeline.
Critical weaknesses

As far as I know, this work has no substantive innovation. Entropy-regularized OT pseudo-labeling, swapped prediction, DEC-style target sharpening, and neighborhood-distribution regularization are established techniques. The paper mainly combines these components, but does not clearly establish a substantial difference from the closest related methods.

I don't see what's interesting about this work. Also, the paper says "extensive experiments" in many places, but with only 4 datasets, that doesn't deserve to be called extensive experiments. I'm not asking the authors to add more experiments, I just think there's some overclaiming here. They could just say "multiple datasets" instead of "extensive experiments." By the way, formulas that are not cited do not need numbering.

Questions for authors
  1. How are X1X^1 and X2X^2 constructed for the four datasets? What is the exact formula for the claimed frequency-domain view?

  2. Are the MLP encoders shared across views or independently parameterized?

  3. How is deg⁡proto(j)\deg_{\mathrm{proto}}(j) computed? What graph or topological structure is associated with a prototype?

  4. What is the update rule for w\mathbf{w}? If b\mathbf{b} is non-uniform, in what sense are the resulting assignments globally balanced?

  5. Is softmax⁡(Lv)\operatorname{softmax}(L^v) used in Eq. (12)? If not, how is the KL divergence defined for unnormalized logits?

  6. How is the hard assignment matrix YY in Eq. (17) obtained and updated? Is it detached from the computational graph?

  7. What pseudo-label generation method is used in the ``w/o OT'' variant?

1 comment
Reviewer_m4yAOpenReview ↗

ICLR Rebuttal: Please Submit Your Response Soon

This is a reminder that the ICLR rebuttal period is ending soon. I would appreciate it if the authors could submit their response as soon as possible so that I can finalize my review.

Reviewer_xzLuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes OT-DDC, a multi-view graph clustering framework. Its core design consists of: constructing two complementary views via Laplacian smoothing and frequency-domain transformation, aligning node representations across views through cross-view consistency learning; introducing an entropy-regularized optimal transport (OT) module to generate globally balanced cross-view soft pseudo-labels; and designing a dual-distribution constraint mechanism that jointly optimizes cluster centroids in both feature space and neighborhood-distribution space. The paper compares against several baselines on multiple datasets.

Critical strengths

S1. The paper addresses meaningful challenges in multi-view graph clustering, including weak cross-view alignment and the mismatch between feature representations and graph topology.

S2. The overall framework is reasonably motivated. Combining globally coordinated pseudo-label generation with feature-space and neighborhood-distribution constraints is an interesting direction.

S3. The paper evaluates the proposed method on multiple graph clustering benchmarks and compares it with a range of graph clustering baselines. The reported results indicate that the method is competitive.

Critical weaknesses

W1. Equation (5) defines L^v as the similarity between nodes and prototypes, and uses it in the KL-divergence loss in Eq. (12). Since the inner product can be negative and is not normalized, it cannot directly serve as a valid probability distribution.

W2. Equation (6) uses degree differences as the topology-aware component of the OT cost. Since degree captures only a coarse aspect of local connectivity, the rationale for the degree-based topology cost could be clarified.

W3. The construction of the two views is insufficiently described, especially for the construction of the frequency-domain view.

Questions for authors

Q1. How is the prototype degree in Eq. (6) computed?

Q2. Why is optimal transport effective in this method?

Reviewer_8Y2JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets multi view graph clustering and aims at addressing the lack of glob al supervision for cross view representation alignment. It also considers the issue of topological neighbor distribution, and introduces entropy regularized optimal transport to generate globally balanced cross view soft pseudo labels for self-supervision. It also incorporates dual distribution constraint to jointly optimize feature space cluster centroids and topology aware neighborhood distribution centers, aligning semantic node representations with graph topological community structures. Extensive experiments including comparison, ablation studies, hyper parameter analysis and t-SNE visualization are reported. Nevertheless, several issues regarding normalization, update mechanisms, metric scales and experimental reproducibility need to be resolved.

Critical strengths

1.The motivation is well grounded. The paper correctly identifies two major limitations in multi view graph clustering: existing methods mostly rely on local supervision without global constraints for cross view alignment; most approaches only regularize node features while ignoring topological neighborhood distributions. 2.The method exhibits certain novelty. Optimal transport is adopted to establish global node to prototype matching and produce soft pseudo labels incorporating both attribute and topological signals, differing from conventional independently generated pseudo labels. The dual distribution constraint jointly regularizes clustering by fusing semantic features and graph topology for clustering optimization. 3.The experimental pipeline is relatively complete. The authors evaluate their approach on standard graph datasets against a broad set of deep graph clustering and contrastive graph clustering baselines. Ablation studies, hyper parameter sensitivity analysis and t SNE visualizations are provided to support model effectiveness.

Critical weaknesses

1.In Equation (7), M_topo lies within range [0,1], while attribute cost M_attr originates from unbounded innerproduct computation. The two terms possess incompatible numerical scales. Fixing alpha_ot=0.5 naively mixes cost terms of different magnitudes without justifications from theory or auxiliary ablation results. 2.Since both views share the same adjacency matrix, why is the topology cost matrix M_topov view-dependent? If the node degrees and prototype-degree definitions are identical across views, M_topov should remain unchanged, making the superscript v redundant and potentially notationally inconsistent. 3.It is unclear why hard assignments Y are used to construct the neighborhood distribution centers instead of using soft cluster assignments.

Questions for authors

Refer to the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.