acceptodds
Withdrawn from ICLR 2027

Consensus-Guided Deep Multi-View Clustering with Dual Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-view clustering aims to leverage complementary information from different data sources to improve clustering performance. However, many existing methods emphasize cross-view consensus in deep representations while underexploiting structural cues retained in shallower features, leaving complementary information underused and making the learned clusters less discriminative. Moreover, they often explore clustering patterns through centroid assignment or alignment, insufficiently modeling inter-sample relationships. Same-cluster samples may remain dispersed while different-cluster samples lie close together, leaving clusters dispersed with unclear boundaries. To address these challenges, we propose COnsensus-guided deep multi-view clustering with dual REpresentation learning (CORE), a framework that combines multi-hierarchical representation learning, semantic invariance, and hard-samples contrastive alignment. Specifically, CORE employs deep and shallow branches to extract abstract semantics and preserve structural cues, respectively, allowing the model to use complementary information when learning discriminative representations. It then constructs a global clustering target from concatenated deep representations to guide semantic invariance across views. Our analysis derives an explicit bound showing when a discriminative view’s distance advantage outweighs conflicts to determine the consensus alignment. Finally, CORE uses consensus-derived pseudo-labels to jointly mine hard positive and negative samples. Contrastive learning then increases the similarity of the selected hard positives and decreases that of the hard negatives, improving separation between clusters. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-View Clustering, Contrastive Learning, Dual Representation Learning

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_kdZqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CORE, a consensus-guided deep multi-view clustering framework with dual representation learning. It introduces two representation branches for each view: a deep branch to capture abstract semantic information and a shallow branch to preserve structural cues. A consensus target constructed from deep representations is used to encourage semantic invariance across views, while a hardness-aware contrastive learning module mines hard positive and negative samples to improve intra-cluster compactness and inter-cluster separation. The authors further provide theoretical analysis of the consensus mechanism and evaluate the proposed framework on several benchmark multi-view clustering datasets, showing improved performance over compared baselines.

Critical strengths
  1. The combination of semantic representation learning, structural representation preservation, and contrastive refinement provides a reasonable approach for improving cluster discrimination.

  2. The paper evaluates the proposed method on multiple benchmark datasets and provides ablation studies, visualization results, and parameter analysis, which offer useful evidence for the effectiveness of the framework.

Critical weaknesses

1. The motivation and problem formulation are not sufficiently differentiated from existing deep multi-view clustering methods.

The paper motivates CORE from the observation that existing methods mainly exploit deep representations while underusing shallow structural information. However, this distinction is not clearly established. Many existing deep MVC methods already explore complementary information, hierarchical representations, reconstruction constraints, graph structures, or contrastive objectives to preserve semantic and structural information.

The paper does not clearly demonstrate that the missing shallow structural information is a fundamental limitation of previous approaches rather than an alternative architectural design choice. In particular, it remains unclear what type of clustering information is uniquely captured by the shallow branch and why existing representation learning strategies cannot preserve such information.

2. The novelty and multi-view-specific advantage of the hard sample contrastive learning component are insufficiently justified.

The proposed HaSE module performs pseudo-label-guided hard positive and hard negative mining followed by contrastive learning. However, this design is closely related to existing contrastive learning paradigms, where difficult positive and negative pairs are selected to improve representation separability.

The paper does not clearly explain what is unique about hard sample mining in the multi-view clustering scenario. For example, it is unclear whether the selected hard samples correspond to view disagreement, unreliable consensus, or other specific multi-view ambiguities. Without such analysis, HaSE appears to be a general contrastive refinement strategy applied to MVC rather than a method addressing a unique challenge in multi-view clustering.

3. The theoretical analysis is not sufficiently connected to the main methodological contributions.

The theoretical results characterize when a discriminative view can dominate the consensus assignment. While the analysis is mathematically interesting, it mainly explains the behavior of the consensus mechanism rather than the necessity or effectiveness of the proposed dual representation learning and hard sample contrastive refinement. The connection between the theoretical results and the claimed advantages of CORE remains unclear.

Questions for authors
  1. Existing MVC methods have extensively studied complementary information, semantic consistency, and structural preservation. What specific limitation of previous methods motivates introducing a separate shallow representation branch? Can the authors provide analysis showing that the shallow representation captures cluster-relevant information that is not already contained in the deep representation?

  2. Several previous MVC methods already employ pseudo labels, cluster assignments, prototypes, or instance relationships for contrastive learning. What is the fundamental difference between HaSE and these existing contrastive MVC approaches? Is the improvement mainly due to harder pair selection, or does HaSE exploit a unique property of multi-view clustering?

  3. Since HaSE relies on consensus-derived pseudo labels, how robust is CORE when the initial consensus representation is inaccurate or when different views provide conflicting clustering evidence? Is there any mechanism preventing incorrect pseudo labels from being reinforced during contrastive learning?

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

CORE combines deep–shallow representation learning, consensus-guided semantic alignment, and pseudo-label-guided hard-example mining for multi-view clustering. Evaluation covers six datasets and nine baselines, with ablation, convergence, and parameter analyses.

Critical strengths
  1. The comparisons include recent methods from 2024–2026, such as BONE and BMvC, and report means and standard deviations.
  2. The paper includes component ablations, parameter analyses, and conditional theoretical results explaining consensus formation.
Critical weaknesses
  1. The structural role of the shallow branch is insufficiently validated. A linear encoder with reconstruction supervision does not itself establish preservation of structural or contour information. CORE-S does not distinguish this benefit from additional parameters and supervision. Parameter-matched controls and training-cost comparisons would strengthen the evidence.
  2. Pseudo-label reliability and hard-example mining need targeted validation. Incorrect pseudo-labels may be amplified by selecting the hardest positives and negatives. Confidence filtering primarily selects anchors, while candidate reliability remains unclear.
  3. Presentation details require clarification. Figure 1 should explain the view definitions and whether encoder parameters are shared. Its (\mathcal L^{CoCA}) notation should match (\mathcal L_{\mathrm{HaSE}}) in the text. Equation punctuation, including Eqs. (1) and (21), should also be checked.
Questions for authors

1.Do existing results show how pseudo-label quality changes during training? Are mined positive and negative candidates also confidence-filtered?

Reviewer_7s6aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a consensus-guided dual representation learning framework. The core contributions are as follows: high-level semantic and low-level structural cues are extracted separately, and a global clustering objective is constructed based on concatenated deep representations; an explicit mathematical bound is theoretically derived showing that highly discriminative views dominate global consensus alignment; and positive and negative samples are dynamically mined based on high-confidence consensus pseudo-labels for contrastive learning, thereby enhancing intra-cluster compactness and inter-cluster separation. Experiments show that CORE is compared with 9 state-of-the-art methods on 6 multi-view benchmark datasets, verifying the superior performance of the model.

Critical strengths

The model adopts two branches, deep and shallow: the deep branch is responsible for extracting high-level abstract semantics and generating reliable global pseudo-labels accordingly, while the shallow branch is responsible for capturing low-level structural cues to supplement complementary information.

Critical weaknesses
  1. The authors conduct deep learning-based multi-view clustering research, but the datasets used have rather small sample sizes. For example, MSRC has only 210 samples and ORL has only 400 samples, while the largest dataset, ALOI, contains 10,800 samples.
  2. The second module clearly overlaps with Section 3.1 of SDMVC (doi: 10.1109/TKDE.2022.3193569): Equations (3), (4), and (5) of the original paper closely correspond respectively to Equations (4), (11), and (12) of SDMVC, while representation concatenation, global K-means, and global soft assignment also correspond respectively to Equations (5), (6), and (10) of SDMVC. Moreover, the claim emphasized in Equations (7)–(10) of the original paper that highly discriminative views dominate the global consensus is not a new observation, since Equations (7)–(9) of SDMVC have already demonstrated, by decomposing the global distance and analyzing the case of one low-discriminative and one high-discriminative view, that this distance advantage can determine the final clustering assignment.
  3. The paper does not clearly specify the optimizer, learning rate, batch size, detailed layer dimensions, training/pretraining schedule, or hardware/software environment. Several hyperparameters introduced in the method also lack clear implementation settings, which makes reproduction difficult.
Questions for authors

Could the authors clarify how Theorem 1 in Eqs. (7)–(10) differs from the phenomenon already established by SDMVC's Eqs. (7)–(9)? Specifically, does Theorem 1 provide a stronger formal guarantee (e.g., explicit bounds or conditions) that SDMVC's two-view/two-cluster analysis does not?

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

Related papers

Discussion (0)

Sign in to comment.