acceptodds
Withdrawn from ICLR 2027

Multi-view Graph Clustering with Partial Information Decomposition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-view graph clustering aims to partition graph elements, including graphs and nodes, by exploring diverse views with distinct semantic topology patterns. Existing methods obtain discriminative representations mainly by integrating multi-view features or modeling consistency across views. However, these strategies assume cross-view information is entirely shared, which ignores other possible types of interaction, degrading the robustness and applicability of clustering. To this end, we propose a multi-view graph clustering framework based on partial information decomposition (PIDGC) that enables communication between multiple graph views in a universal space. Instead of enforcing restrictions within or across views, we achieve alignment of multiple graph views by optimizing the mutual information between original views and corresponding augmented counterparts, exploring redundant interactions between different views. We provide a theoretical illustration to demonstrate that mutual information can naturally explore diverse cross-view information via partial information decomposition, improving the discrimination of representations and clustering performance. Extensive experiments on multifarious datasets demonstrate the effectiveness and advancement of PIDGC compared with state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view graph clustering, partial information decomposition, graph contrastive learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7W3tOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

For multi-view graphs, the authors propose to learn node representations via 1) contrastive learning between latent features learned from an original graph view and its masked version; 2) k-means clustering loss on the latent feature aggregated from all views. The paper presents promising evaluation results with an extensive set of baseline methods.

Critical strengths

None

Critical weaknesses
  1. The third paragraph of the introduction section claims that PID, or the three types of information, is used to "further illustrate and interpret the interaction between different views". However, in the formulation of the L_1 and L_2 loss functions, the contrast is only between the learned features H_i of view i and two sets of learned features H_i' and H_i'' of two masked versions of view i. All the contrastive comparisons are between representations of the same view, without any cross-view interaction.
  2. The proof of Theorem 1 in appendix doesn't seem to establish I(Hθ∗,Y)=I(X,Y)I(H_{\theta^*}, Y)=I(X,Y).
  3. Section 4.1 mentions data flow "X-> X'->H_\theta" and "H_\theta'->X->H_\theta'" without justification.
  4. Section 4.2 simply mentions "GNNs as basic encoders" without explaining the model actually used.
  5. Essentially, the proposed method uses InfoNCE (an earlier reference for this approach should be cited instead of the 2026 paper) to contrast View i's learned features against those learned from its masked version, and a clustering loss using weighted aggregation of single-view features. The contrastive learning part has highly similar precedents, such as "A Two-View Hierarchical Contrastive Learning-Driven Method for Community Detection" (2026, Eq 19-22), and "Graph Contrastive Learning with Adaptive Augmentation" (2020). The clustering loss has also been seen in previous work such as "Homophily-Related: Adaptive Hybrid Graph Filter for Multi-View Graph Clustering" (2024).
  6. The evaluation includes a task where clustering is done on graphs. It's unclear how this is related to the multi-view graph setting of this paper.
Questions for authors
  1. How does the use of PID in this paper address "possible alternatives of the relationship between different views"?
  2. How is "data flow" defined, and how are these two flows in Section 4.1 derived?
Reviewer_jaUHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PIDGC, a multi-view graph clustering framework inspired by partial information decomposition. The method aims to capture redundant, unique, and synergistic information across graph views through contrastive learning. It combines view-specific graph encoders, feature augmentations, adaptive representation fusion, and clustering refinement.

Experiments on six node-level and six graph-level datasets show consistent improvements over existing baselines. The authors also provide theoretical analysis and ablation studies to support the proposed design.

Critical strengths
  1. The motivation is interesting. Considering redundancy, uniqueness, and synergy provides a useful perspective beyond conventional cross-view consistency learning.

  2. The experimental results are strong. PIDGC performs well across both node-level and graph-level clustering tasks, and the ablation studies provide evidence that the proposed loss components contribute to performance.

Critical weaknesses
  1. The theoretical justification needs to be strengthened. Lemma 2 and Theorem 1 claim that maximizing mutual information between augmented representations preserves all task-relevant information. However, the equality conditions are not sufficiently justified. In particular, Assumption 2 does not necessarily guarantee that the augmented representations retain the information relevant to the clustering task. This weakens the main theoretical contribution.

  2. The connection between the proposed objective and PID is not fully convincing. In Eq. (7), the contrastive terms mainly operate on different augmentations of the same view. It is unclear how this objective explicitly distinguishes redundancy, uniqueness, and synergy across views. The performance gains are promising, but they do not directly establish that these information components are successfully captured.

Questions for authors
  1. Could the authors clarify the conditions under which the equality in Lemma 2 and Theorem 1 holds? In particular, why does Assumption 2 guarantee preservation of task-relevant information?

  2. Which part of Eq. (7) specifically captures synergistic information between different graph views? Could the authors explain how the proposed objective distinguishes synergy from ordinary within-view contrastive learning?

Reviewer_6o6UOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that multi-view graph clustering (MVGC) methods implicitly assume that all task-relevant information is shared across views. It proposes PIDGC, which uses Partial Information Decomposition (PID) to motivate InfoNCE losses between original and augmented representations, intended to capture redundancy, uniqueness and synergy. It reports state-of-the-art results on node-level and graph-level clustering.

Critical strengths
  • The motivation is relevant and timely: relaxing the multi-view redundancy assumption is an important question for MVGC.
  • The framework is simple, modular and unified for node-level and graph-level clustering.
  • The evaluation is broad: 12 datasets, recent baselines (2025–2026), ablations, and standard deviations over 5 seeds.
Critical weaknesses
  • The theory contradicts the motivation. Assumption 1 (Sec. 3) is precisely the redundancy assumption the paper claims to relax. The loss (Eqs. 6–7) contains no cross-view term, so synergy, which requires the joint (X1, X2), cannot be captured. The views interact only through the weighted sum in Eq. (5).

  • The proofs are invalid or incomplete.

    • Lemma 1 has no proof.
    • The Markov chain in Eq. (2) is ill-formed.
    • Lemma 2 relies on an identity map outside the hypothesis class.
    • Theorem 1 is internally inconsistent: it gives I(H*, Y) = I(X, Y) and I(H’_i, Y) = R + U_i for the same encoder.
    • The PID variables (X1, X2) are conflated with the shared features X (Sec. 4.1).
  • Some reported numbers are mathematically impossible, and the protocol is inconsistent.

    • Hungarian-matched ACC cannot fall below 1/k, yet ENZYMES (k = 6) shows 11.8% (Table 3) and 15.6% (Table 2).
    • Sec. 5.1 names AUROC/AUPRC as metrics, while all tables report NMI, ARI and ACC.
    • Several baselines show ±0.0 standard deviation over 5 seeds.
    • It is unclear whether baselines were re-run or copied.
  • Hyperparameters may be selected with labels, and the multi-view setting is largely artificial.

    • Augmentations, the similarity measure and the hyperparameters are chosen by ACC (Tables 5–6, Fig. 2), which is label leakage.
    • Most datasets are single-view with a synthetic kNN second view, and DBLP’s three relations are reduced to two.
  • Formal errors and inconsistencies.

    • Eq. (8) omits the square of q_ik, and Eq. (9) reverses the DEC objective.
    • The pseudo-labels ŷ are defined but never used.
    • The method is named “MVGCPI” in Fig. 1.
    • The efficiency claims (Sec. 5.2) contradict Appendix H.

Overall, the paper’s claims are supported neither by its theory nor by a reliable experimental protocol. The closest prior work, FactorCL (Liang et al., NeurIPS 2023), is not cited, although the paper closely mirrors its augmentation assumption and its redundancy/uniqueness/synergy framing; the “first attempt” claim in Sec. 1 is therefore overstated. The language is understandable but often imprecise, and the notation contains several errors. The reference list leans heavily on 2025–2026 papers: InfoNCE should be credited to van den Oord et al. (2018) and DEC to Xie et al. (ICML 2016), Zhang et al. 2026c and 2026d are duplicates, and relevant work such as MVGRL (Hassani & Khasahmadi, ICML 2020), O2MAC (Fan et al., WWW 2020), MCGC (Pan & Kang, NeurIPS 2021), Tsai et al. (ICLR 2021) and Federici et al. (ICLR 2020) is missing.

Questions for authors

Suggestions

  • Reformulate the problem without Assumption 1, add an explicit cross-view objective (e.g., conditional MI terms as in FactorCL), and correct the proofs.
  • Re-run all baselines under a single label-free protocol (internal validity indices) with released code, and report metrics consistently.
  • Evaluate on natively multi-view benchmarks with three or more views, plus a synthetic dataset with controlled R/U/S.

Questions

  • Which loss term involves H₁ and H₂ jointly?
  • How can ACC fall below 1/k?
  • Were the baseline numbers reproduced or copied?

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

Related papers

Discussion (0)

Sign in to comment.