acceptodds
Withdrawn from ICLR 2027

SEMANTICS-GUIDED DISENTANGLED MULTI-VIEW CLUSTERING

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing deep multi-view clustering methods commonly adopt dual-branch paradigms to extract and fuse view-shared and view-specific features, yet they suffer from inherent structural limitations. Direct aggregation of these features introduces view-irrelevant noise and underutilizes cross-view complementary information, while partitioning features by view origins does not align with clustering objectives, due to redundant view-shared features and underexploited discriminative cues in view-specific features. To address these issues, we propose a Semantics-Guided Disentangled paradigm for Multi-View Clustering (SG-DMVC). Specifically, we explicitly decompose the input of each view into discriminative cluster-relevant features and task-agnostic cluster-irrelevant features. The view-wise cluster-relevant features are aggregated via log-space probability consensus to yield robust cross-view shared semantics, which are further regularized by dual-level topological alignment to preserve global manifold distribution consistency and refine local sample-pair structures. Meanwhile, covariance decorrelation regularization is imposed to enforce second-order decorrelation between cluster-relevant and cluster-irrelevant features, effectively eliminating irrelevant feature interference and latent information leakage. Extensive experiments on both complete and incomplete multi-view datasets show that the proposed SG-DMVC yields superior clustering performance and strong robustness against view missing compared with state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view clustering, disentangled representation, log-space probability consensus, topological alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1feKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the SG-DMVC framework that explicitly decomposes multi-view inputs into cluster-relevant and cluster-irrelevant features, departing from the traditional paradigm. The model utilizes a log-space probability consensus mechanism to extract cross-view semantics, and a dual-level topological alignment to maintain manifold consistency. Furthermore, it introduces covariance decorrelation regularization to eliminate interference from irrelevant features.

Critical strengths
  1. Redefining disentanglement as cluster-relevant vs. irrelevant offers an inspiring and logically sound approach that directly serves the downstream clustering objective.

  2. Theorems 1 and 2 offer solid mathematical backing for the log-space probability consensus based on mutual information bounds.

Critical weaknesses
  1. The framework heavily relies on existing techniques (e.g., KL-divergence, KNN contrastive learning), reading more like a combination of standard methods than a breakthrough algorithmic innovation.

  2. The analysis of "cluster-irrelevant" features is superficial. Scattered t-SNE visualizations (Figure 3) do not guarantee these features capture view-specific semantics; they may merely be collapsed noise, making the method seemingly indistinguishable from a standard denoising autoencoder.

  3. The framework's multiple loss components (L_rec, L_align, L_topo, L_orth) require numerous hyperparameters. Tuning these systematically in a completely unsupervised scenario remains a major practical hurdle, regardless of the provided sensitivity analysis.

  4. Over-reliance on raw-feature KNN graphs for local alignment makes the model vulnerable; poor-quality raw features could introduce misleading relational biases and trigger severe performance collapse.

Questions for authors
  1. Can you provide quantitative experiments beyond t-SNE to prove UvU^v captures meaningful view-specific semantics instead of collapsing into random noise?

  2. Given its reliance on raw features, how severely would SG-DMVC degrade if artificial noise were injected to corrupt the initial KNN graph?

  3. How should the optimal values for α\alpha, β\beta, and λ\lambda be systematically determined for a new dataset in a strictly unsupervised scenario?

Reviewer_Efz1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

SG-DMVC proposes a semantics-guided disentanglement framework for deep multi-view clustering. It decomposes each view into cluster-relevant and cluster-irrelevant representations. The relevant representations are aggregated through log-space probability consensus and refined by global and local topological alignment. Covariance decorrelation and reconstruction losses further constrain the latent spaces. Experiments evaluate the method on complete and incomplete multi-view datasets.

Critical strengths
  1. The paper identifies the limitations of conventional shared-specific disentanglement and proposes a clustering-oriented semantic decomposition.
  2. The paper studies information bounds, consensus estimation error, and topological consistency, providing a theoretical perspective on the proposed design.
  3. The evaluation includes multiple data domains, different numbers of views, incomplete-view settings, and loss-component ablations.
Critical weaknesses
  1. The consensus operation does not clearly distinguish representations from probabilities, while the global alignment objective differs between the main text and supplementary theoretical formulation.
  2. The construction of positive and negative neighborhoods and the confidence gate is unclear.
  3. The method equations do not explicitly define masking and normalization for missing observations, while the supplementary material restricts the theoretical analysis to complete-view inputs.
  4. The manuscript contains inconsistencies in reporting and interpretation.
    • The main text uses λ\lambda for the covariance term, while the supplementary implementation details call the covariance weight γ\gamma.The main text and supplementary material use λm\lambda_m, α\alpha, β\beta, λ\lambda, and γ\gamma with different roles across equations and descriptions.
    • The statement that the full model is best across all Table 4 datasets is contradicted by Wikipedia PUR, where Lrec+LtopoL_{\mathrm{rec}}+L_{\mathrm{topo}} reaches 64.07 while the full model reaches 63.78.
    • The implementation section contains a faulty cross-reference “fully-connected encoder-decoder backbones (?)”.
Questions for authors

Please address the questions raised in the Critical Weaknesses section.

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

This paper identifies a drawback of existing mainstream disentanglement methods: simply concatenating view-specific and view-shared features introduces irrelevant noise and fails to fully exploit cross-view complementary information. To address this issue, the authors propose a semantics-guided disentangled paradigm for multi-view clustering. Instead of partitioning latent variables according to view sources, the proposed approach decomposes each view input into two components: cluster-relevant latent features and cluster-irrelevant latent features. The authors conduct comprehensive experiments on eight commonly used datasets and achieve promising results.

Critical strengths

(1) This work further refines the existing disentanglement paradigm. It points out that noise exists in view-shared features and that view-specific features are underutilized. The paper explicitly decomposes the input of each view into discriminative features relevant to clustering and task-irrelevant features, offering a new insight for multi-view clustering. (2) The proposed SG-DMVC in this manuscript is a relatively original technique. Besides, the writing quality of this paper is relatively high, with almost no obvious typos or grammar errors.

Critical weaknesses

(1) Equation (6) contains four loss terms and three balancing hyperparameters. In addition, there are other hyperparameters such as the temperature coefficient and neighbor count. In my opinion, this makes the proposed SG-DMVC difficult to tune, limiting its practical deployment in real-world scenarios. (2) The paper claims strong robustness against view missing, yet quantitative experiments under different missing ratios are absent. Detailed comparisons with dedicated missing-view methods across various missing rates are not provided, resulting in insufficient evidence to justify its robustness. (3) Except for STL-10 and NUSWIDEOBJ, the remaining datasets are too small to guarantee convincing and solid experimental results.

Questions for authors

(1) The authors state that the experiments adequately verify the basic motivation of the proposed SG-DMVC model, while I hold a different opinion. In my view, the experiments only demonstrate that SG-DMVC achieves superior clustering performance compared with other baselines, but cannot validate the core motivation. To verify that SG-DMVC can filter irrelevant noise and fully utilize view-specific information, visualizations or other evidence are required. For instance, visualizing samples from the Digit dataset after training to demonstrate that noise is indeed eliminated. I expect the authors to address this concern. (2) The theorem relies on idealized assumptions including manifold perturbation models and bi-Lipschitz mappings. The paper acknowledges that real-world data can hardly satisfy these assumptions strictly. The theory mainly provides qualitative inspiration, and further analysis of model behavior when assumptions are violated is missing, which weakens the practical interpretability of the theoretical part. I hope the authors provide explanations on this issue. (3) Other questions can be found in Critical Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.