acceptodds
Under review as a conference paper at ICLR 2027

Rethinking View Consistency: Consensus-Oriented Representation Learning for Multi-View Clustering

Abstract

Multi-view clustering aims to integrate heterogeneous observations from multiple views to learn a discriminative consensus representation for clustering. Existing methods often place cross-view consistency at the center of representation learning, either by directly aligning different views or by using a global target to constrain individual-view representations. While effective when clustering-relevant information is redundantly distributed across views, such consistency-oriented strategies can become restrictive when discriminative evidence is available only in a subset of views. Enforcing cross-view consistency may weaken view-specific informative structures, while global-to-local constraints may force individual views to reproduce semantics beyond their own observational scope. In this work, we rethink the role of view consistency and propose a consensus-oriented representation learning method for multi-view clustering. Instead of explicitly driving individual views toward cross-view consistency, we preserve their intrinsic relational structures and allow them to contribute differently to a jointly learned consensus. Direct clustering supervision is imposed only on the consensus representation, allowing each view to be optimized according to its contribution to the global clustering structure rather than being forced to match the complete global target. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method effectively preserves view relations and yields more discriminative clustering results. Code will be available upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view, clustering, consensus, consistency

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_u5NeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript challenges the popular paradigm of consistency learning in multi-view clustering. After summarizing existing consistency-based methods into two categories (strengthening cross-view consistency, and obtaining a global target for supervision), the authors provide an analysis of the effectiveness and limitations of consistency learning, thus proposing a consensus-oriented multi-view representation learning framework. The framework consists of three parts: relation preservation, self-attention and consensus-only clustering. Extensive experiments have shown the effectiveness of this framework.

Critical strengths

1.The method is well-designed and achieves strong performance on a wide range of benchmarks.

2.The authors have made efforts to give a theoretical analysis on the limitations of consistency learning.

Critical weaknesses

1.The definition of the word “consensus learning” is not clear, therefore the central contribution remains insufficiently defined. In my view, this study mainly works on maintaining complementary information across different views, which falls exactly into the perspective of complementary learning of MVC. However, the authors do not clarify on their relation with and differences from existing complementarity-aware MVC literature. Indeed, the authors mainly review consistency-based methods, while this work is substantially more related to complementary learning.

2.The analysis is rather simple.

Questions for authors

1.What is the definition of “consensus learning”? In what sense is your study different from existent complementary learning in MVC?

2.The quality of Figure 1 is not good. It is not immediately clear that the three images represent different views. Also, the supervision relation of the right part is not so clear.

3.The analysis seems to be an important part of your work as it is listed in your main contributions. However, it is altogether simple. Despite your complicated mathematical presentation, you do not present anything other than the two paradigms of consistency-based learning and their failure modes.

4.In the last paragraph of pg.4, you mention two principles: (1) preserve each view’s local relational structures during representation learning; and (2) apply direct clustering supervision only at the consensus level. In Table 2, you mainly ablate different consensus formation and relation preservation. There is insufficient direct empirical evidence on the correctness of these two principles.

8 comments

1, We thank the reviewer for the question.

First, we define Consensus Learning as a multi-view representation learning paradigm that integrates clustering-relevant evidence across views at the consensus level and directly learns the clustering structure. Its central idea is that each view provides structural evidence consistent with its own observations, while the consensus representation integrates this evidence and carries the final clustering semantics. Therefore, an individual view does not need to independently possess complete clustering-discriminative information; the clustering structure is determined jointly by the integrated multi-view evidence.

Second, our distinction from existing complementary learning methods lies in how inter-view information relationships and their optimization objectives are treated. Existing complementary learning methods primarily aim to explore and exploit differences across views to obtain richer sample representations and clustering-discriminative information. However, differences across views are not necessarily relevant to clustering and may also contain information unrelated to the final clustering structure. In contrast, Consensus Learning does not prescribe whether views should be consistent or complementary, but treats both as manifestations of the relationships among multi-view evidence during the clustering process.

Specifically, in our method, the relational preservation loss constrains the local structures within each view, with the aim of preserving the fidelity of each view to its own observations, rather than prescribing the information relationships between views. Accordingly, we do not explicitly constrain inter-view consistency or complementarity. We then construct the consensus representation through sample-wise adaptive cross-view interactions and apply clustering supervision only at the consensus level. This allows the contributions of different views to the final clustering structure, as well as their representation learning, to be guided by the clustering objective. Consequently, both redundant and complementary clustering-relevant evidence can contribute to consensus formation without requiring their information relationships to be specified in advance. We have added the relevant discussion to Section 2.2 of the revised manuscript.

2, We thank the reviewer for the suggestion. We have redrawn Figure 1 to clearly indicate that the three images represent different views, and revised the schematic on the right to better illustrate the relationship between the consensus representation and clustering supervision. These changes are reflected in Figure 1 of the revised manuscript.

3, We thank the reviewer for the comment. We would like to further clarify the full scope and main findings of our theoretical analysis. First, due to the page limit, Section 3.1 provides a concise analysis of the relationship between consistency learning and consensus learning, primarily based on conditional mutual information. Specifically, we use the conditional mutual information (\Gamma_{ij}^{v}) to characterize the additional discriminative information that different views provide for the clustering relationship between a sample pair. This analysis identifies when consistency learning can approximate consensus learning: when clustering-relevant information is largely redundant across views. We then present an abstract formulation of consensus learning and establish two basic principles for consensus-oriented representation learning. Second, the underlying mechanisms and limitations of the two consistency-oriented paradigms are further examined in Appendix A.

Appendix A.1 examines how cross-view consistency constraints affect discriminative structures. We first consider a consistency constraint based on squared Euclidean distance and reformulate it as a common-center optimization problem. This explains why discriminative information supported by only a subset of views may be weakened. We then analyze consensus formation using a general consensus function, showing how the clustering objective can guide the integration of multi-view information without requiring discriminative information to be present in every view.

Appendix A.2 analyzes the information mismatch in global-to-local supervision. Specifically, we rewrite the conditional mutual information as a difference between conditional entropies and explain, from an information-theoretic perspective, that when other views contain clustering-relevant information unavailable from an individual view, the global objective may involve clustering semantics beyond what that view can capture independently. Directly requiring local predictions to match the global objective may therefore lead to supervision mismatch. Furthermore, through the gradient propagation of the consensus-level clustering objective, we explain how each view participates in clustering optimization through its contribution to the consensus representation, without having to independently reproduce the complete global clustering semantics.

Overall, our analysis goes beyond summarizing the two consistency-oriented paradigms and their potential failure cases. It characterizes the relationship between consistency learning and consensus learning, including the conditions under which consistency learning can approximate consensus learning. More importantly, these analyses explain why consensus-oriented representation learning should preserve the local relational structure of each view and apply direct clustering supervision at the consensus level.

4, We thank the reviewer for the comment. We verify the two principles by corresponding experiments.

First, the principle of preserving each view's local relational structures during representation learning is implemented through the relational preservation loss. The ablation study on this constraint in the original Ablation Study subsection provides a direct empirical evaluation of this principle. The results show that incorporating relational preservation consistently improves performance under different consensus formation strategies, indicating that maintaining the local relational structures of individual views benefits multi-view clustering.

Second, to evaluate the necessity of applying clustering supervision only at the consensus level, we considered three supervision strategies: consensus-only, view-only, and joint supervision. Joint supervision was examined in the original Empirical Analysis of Consistency-Oriented Learning subsection by varying the balance between the supervision terms. The results showed that joint supervision did not lead to better clustering performance. In addition, we conducted a new ablation study with supervision applied only to view representations. By keeping all other settings unchanged and varying only the location of clustering supervision, we directly compared consensus-only and view-only supervision. As shown in the new Table 3, consensus-only supervision outperforms view-only supervision on all evaluation metrics across the three datasets. These results are consistent with the theoretical analysis in Section 3.1 and Appendix A.2. When other views provide additional clustering-discriminative information that is absent from a particular view, requiring each view to match the global clustering target may lead to a global-to-local supervision mismatch. In contrast, consensus-level supervision first integrates discriminative information from different views and then optimizes the clustering structure at the consensus level, without requiring each view to independently reproduce the complete global clustering semantics. The new supervision-position ablation study and its results have been added to Table 3 of the revised manuscript.

SupervisionAnimalAnimalAnimalAnimalAWA-7viewAWA-7viewAWA-7viewAWA-7viewXmediaXmediaXmedia
accnmipurariaccnmipurariaccnmipurari
View-only0.5920.6770.6510.4910.2030.3330.2390.1030.1360.3840.1560.065
Consensus-only0.6480.7220.6810.560.5580.6350.5950.430.9190.9210.9190.85

We hope that our clarifications, additional experiments, and corresponding revisions adequately address your concerns. We would appreciate your reconsideration of the overall assessment in light of these responses.

Reviewer_1jkyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a consensus-oriented representation learning framework for multi-view clustering. Instead of enforcing cross-view consistency, it preserves view-specific neighborhood relations and integrates representations through self-attention with a learnable consensus token. Clustering supervision is applied only to the consensus representation using refined pseudo-labels. The paper analyzes redundant and complementary information across views and evaluates the method on twelve benchmarks.

Critical strengths
  • The paper provides an interesting perspective on consistency-oriented and consensus-oriented multi-view clustering. Avoiding global-to-local supervision when individual views contain incomplete information is reasonably motivated.

  • The framework is straightforward and achieves competitive results on twelve benchmarks. The ablation study provides some evidence for the effectiveness of relational preservation across different fusion strategies.

Critical weaknesses
  • The proposed framework combines established components, including relation preservation, self-attention fusion, and consensus-level clustering supervision. Its methodological novelty over existing multi-view clustering approaches is not sufficiently demonstrated.

  • Fig. 3 mainly examines the effects of additional consistency constraints, while Table 2 evaluates relation preservation and fusion strategies. Neither directly demonstrates that the proposed method better preserves or exploits complementary discriminative information across views.

  • The experiments lack computational efficiency comparisons, convergence analysis, multi-run stability results, and robustness evaluations under noisy or degraded views. These omissions limit the assessment of the method's efficiency, stability, and practical applicability.

  • In Eqs. (6)–(7), independently constructed neighborhood distributions may have different supports, potentially resulting in an infinite KL divergence. Furthermore, preserving input-space neighborhoods may reinforce irrelevant or noisy relations that are inconsistent with clustering semantics.

  • The geometric assumptions in Appendix A.1 do not directly follow from the conditional mutual information criterion in Eq. (2). Moreover, the attenuation derived in Eq. (21) is specific to averaging and does not establish clustering information loss or justify the proposed attention-based consensus formation.

Questions for authors
  • What is the key methodological novelty beyond combining existing components, and how is the benefit of consensus-only supervision isolated?

  • Can the authors provide controlled experiments demonstrating improved utilization of complementary information distributed across different views?

  • Can the authors provide convergence, computational efficiency, multi-run stability, and noisy-view robustness evaluations?

  • How is the KL divergence handled when neighborhood supports differ, and how are misleading input-space relations addressed?

  • How does the information-theoretic criterion in Eq. (2) support the geometric analysis and the proposed attention-based fusion?

5 comments

1, Methodological novelty: Thank you for the question. Our key methodological contribution is a sample-wise consensus-oriented multi-view representation learning framework, rather than a new network component. The framework adaptively integrates information from different views for each sample to construct a sample-wise consensus representation. It preserves local structures through intra-view relational constraints while applying clustering supervision only at the consensus level. This design differs from methods that directly enforce consistency across view representations or cluster assignments, as well as global-to-local supervision approaches that transfer global clustering semantics to individual views. Therefore, our focus is on how to preserve the intrinsic structure of each view, adaptively integrate multi-view evidence, and determine the final clustering structure at the consensus level. Importantly, the framework does not rely on any specific network component; existing components are simply concrete choices for implementing the proposed design principles, rather than the source of the framework's novelty.

Validation of the benefits of consensus-only supervision: Thank you for the question. To evaluate the benefits of applying clustering supervision only at the consensus level, we considered three supervision strategies: consensus-only, view-only, and joint supervision. Joint supervision was already examined in the original Empirical Analysis of Consistency-Oriented Learning subsection by varying the balance between the supervision terms. The results showed that adding view-level clustering supervision did not yield better performance than consensus-level supervision alone. In addition, we conducted a new ablation study with supervision applied only to view representations. By keeping all other settings unchanged and varying only the location of clustering supervision, we directly compared the two supervision strategies. As shown in the table below, consensus-only supervision outperforms view-only supervision on all evaluation metrics across the three datasets. These results are consistent with the theoretical analysis in Section 3.1 and Appendix A.2. When other views provide additional clustering-relevant information that is absent from a particular view, the global clustering target formed from multiple views may contain information beyond that view's own observations. Directly requiring each view to match this global target may therefore lead to a global-to-local supervision mismatch. In contrast, consensus-level supervision first integrates discriminative information from different views and then optimizes the clustering structure at the consensus level, without requiring each view to independently reproduce the complete global clustering semantics. Taken together, the joint and individual supervision experiments support the benefit of consensus-level clustering supervision, showing that concentrating clustering supervision on the consensus representation, rather than directly imposing it on individual views, is more effective in integrating discriminative information across views. The new supervision-position ablation and its results have been added to Table 3 of the revised manuscript.

SupervisionanimalanimalanimalanimalAWA-7viewAWA-7viewAWA-7viewAWA-7viewxmediaxmediaxmediaxmedia
accnmipurariaccnmipurariaccnmipurari
View-only0.5920.6770.6510.4910.2030.3330.2390.1030.1360.3840.1560.065
Consensus-only0.6480.7220.6810.560.5580.6350.5950.430.9190.9210.9190.85

2, We thank the reviewer for this suggestion. To evaluate how well different methods utilize complementary information across views, we conducted a controlled synthetic study by varying the distribution of discriminative information across views and examining the clustering performance under three settings. The results show that all methods achieve perfect clustering performance in the Fully Shared setting. As the shared information across views decreases, CVCL and AWDCMVC, representing direct cross-view consistency constraints and global-to-local supervision, respectively, exhibit progressively declining performance. In contrast, DDMVC, which jointly learns consistency and diversity, shows an initial decline followed by relative stabilization, possibly benefiting from its diversity constraint that helps preserve view-specific information. Our method, however, maintains consistently strong clustering performance across all three settings. These results demonstrate its ability to effectively integrate complementary discriminative information distributed across views and adapt to different cross-view information distributions. The experimental settings, results, and analysis have been added to Section 4.2 of the revised manuscript. In addition, upon acceptance of the paper, we will release the complete data generation code and the generated synthetic datasets in our public code repository to ensure reproducibility.

MethodFully SharedFully SharedFully SharedFully SharedPartially SharedPartially SharedPartially SharedPartially SharedDisjointDisjointDisjointDisjoint
accnmipurariaccnmipurariaccnmipurari
ours111111111111
AWDC-MVC11110.7030.7810.7030.5930.5460.6960.5690.443
DDMVC11110.7060.7600.7190.5690.6540.7770.6540.563
CVCL11110.5010.6070.5150.3510.2980.2610.3080.112

3, We thank the reviewer for this suggestion. To address these concerns, we added corresponding evaluations to the appendix of the revised manuscript. Section B.3 presents the training convergence curves, demonstrating the good convergence behavior of our method. Section B.4 compares the computational efficiency of different methods, showing that our method has comparable efficiency to mainstream approaches. Section B.5 evaluates clustering stability across multiple random seeds, while Section B.6 examines performance under different noise levels and demonstrates the robustness of our method to noisy views.

4, We thank the reviewer for raising this issue.

First, regarding the computation of the KL divergence, the phrase “in the same manner” in the original manuscript was insufficiently precise and may have led readers to assume that both graph structures use the same KNN neighborhoods. In fact, they use the same cosine similarity measure and temperature scaling but differ in their normalization domains: we construct a KNN graph only in the raw feature space (\mathbf{X}), whereas for the view embeddings (\mathbf{Z}), similarities are computed and normalized over all non-self samples. Consequently, the support of the raw-feature relation distribution is a subset of that of the embedding relation distribution. Given that we use (D_{\mathrm{KL}} \left( R(\mathbf{X}{\mathcal B}^{v}) ,\middle|, R(\mathbf{Z}{\mathcal B}^{v}) \right)), the KL divergence is well-defined. We have clarified this construction in the revised manuscript.

Second, we acknowledge the potential risk raised by the reviewer: local similarities in the raw feature space may not accurately reflect the true clustering relationships and may even introduce misleading neighborhoods. We address this issue at two levels. Within each view, the neighborhood size (k) determines the local neighborhood used for relation preservation, while the trade-off parameter (\lambda) controls its influence on representation learning. Across views, our sample-wise adaptive consensus formation computes attention weights based on the input-dependent interactions among different view representations of the same sample. These weights dynamically adjust each view's contribution to the final consensus representation, allowing different samples to exhibit different patterns of cross-view interaction. Thus, the model can still integrate discriminative evidence from other views.

5, We thank the reviewer for the comment.

First, we clarify that Eq. (2) characterizes the redundancy of clustering information across views from an information-theoretic perspective, while the geometric analysis in Appendix A.1 provides an intuitive explanation from the representation-space perspective. The two are not directly connected through a mathematical derivation. Specifically, when Γijv≈0\Gamma_{ij}^{v}\approx 0 for every view (v), the other views provide little additional clustering information. Conceptually, this corresponds to the case in Appendix A.1 where different views repeatedly support the same pairwise discriminative structure, which can be well preserved by consistency constraints. In contrast, when Γijv>0\Gamma_{ij}^{v}>0, the other views still contain additional clustering-relevant information. This corresponds to cases where certain discriminative structures are supported by only a subset of views, and consistency constraints may weaken these view-specific discriminative differences. We have clarified this point in the revised manuscript.

Regarding consensus formation, Eq. (2) does not uniquely determine a particular fusion module. Rather, it indicates that discriminative information for different sample pairs may be distributed across different views. Therefore, consensus formation should adaptively integrate the clustering evidence provided by different views according to the multi-view observations of each sample. Based on this insight, we employ Self-Attention to enable sample-wise adaptive interactions among the view representations of the same sample, allowing the consensus representation to be formed dynamically based on the information provided by each view. We also considered and compared alternative fusion strategies, including Concat+MLP and Mean Pooling. As shown in the Ablation Study subsection of the original manuscript, Self-Attention achieves balanced overall performance across different datasets, supporting its effectiveness as a consensus formation strategy.

We hope that our clarifications, additional experiments, and corresponding revisions adequately address your concerns. We would appreciate your reconsideration of the overall assessment in light of these responses.

Reviewer_AhWoOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper revisits the role of view consistency in multi-view clustering. The authors argue that forcing different views to learn consistent representations may suppress view-specific information that is important for clustering.

They propose a framework that preserves local structures within each view, combines view representations using self-attention, and applies clustering supervision only to the resulting consensus representation.

Experiments on 12 benchmark datasets demonstrate the effectiveness of the proposed method.

Critical strengths

The method achieves strong clustering results across multiple datasets, with consistently high scores on several benchmarks. As can be seen from the extensive Table 1, the results are indeed strong. The paper also includes some empirical analyses of consistency constraints and global-to-local supervision, providing additional insights.

The paper is generally well-written and easy to follow. The motivation is clearly explained, and the proposed framework is presented in a structured way that makes the main ideas accessible.

Critical weaknesses

The main idea is reasonable, but its novelty is not sufficiently established. Prior work, particularly "Reconsidering Representation Alignment for Multi-View Clustering" (CVPR 2021), has already identified the limitations of strict cross-view alignment and explored alternatives. The paper should more clearly explain what new insights its consensus-oriented formulation provides beyond these existing findings.

The theoretical analysis is not sufficiently strong to support the broader claims. The authors use conditional mutual information to characterize complementary information across views, and illustrate how consistency constraints may weaken view-specific information through a simplified averaging example. However, this analysis does not establish that consistency learning generally harms clustering performance, nor does it theoretically justify why the proposed approach should perform better under realistic settings. As a result, the analysis provides an intuitive motivation for the method, but falls short of establishing the general claims made in the paper.

Questions for authors
  1. How does the proposed approach differ fundamentally from prior work such as Reconsidering Representation Alignment for Multi-View Clustering?

  2. The theoretical analysis uses a simplified averaging example to illustrate how consistency constraints may weaken view-specific information. Under what conditions does this conclusion hold more generally?

  3. Can the authors conduct a controlled synthetic study rather than relying on existing benchmark datasets? For example, construct a 16-cluster dataset using four binary latent factors and two views, with three settings: (1) both views contain all four factors; (2) the views share two factors and each contains one view-specific factor; and (3) the views contain disjoint pairs of factors. The sample size, feature dimensionality, noise distribution, and total signal energy per view should be controlled, with ground-truth labels used only for evaluation. Comparing the proposed method against representative cross-view alignment and global-to-local supervision baselines across these settings would directly test whether consistency-based methods degrade as cross-view information becomes more complementary, and whether the proposed method is more robust in such cases.

4 comments

1, Thank you for this question. Both our work and "Reconsidering Representation Alignment" (CVPR 2021) investigate how information from different views affects the final clustering structure, but they differ in both theoretical analysis and methodological design.

Theoretical analysis: "Reconsidering Representation Alignment" primarily examines view-level class separability, analyzing the discriminative capability of different views and how representation alignment may impair the resulting clustering structure. However, real-world multi-view data may exhibit more complex information distributions: the discriminative capability of a view can vary across individual sample pairs rather than being a fixed global property. For example, in hyperspectral images, different spectral bands may have different discriminative power for different land-cover categories. Thus, view-level class separability alone cannot fully characterize the redundancy and complementarity of information across views for individual sample pairs. To address this, we introduce conditional mutual information Γijv\Gamma_{ij}^{v} to characterize the additional clustering-relevant information provided by the remaining views for a specific sample pair. This sample level analysis allows us to identify the conditions under which cross-view consistency can approximate clustering consensus, and explain why the two are not necessarily equivalent.

Methodological: To address the potential drawbacks of indiscriminate representation alignment, Reconsidering Representation Alignment proposed SiMVC and CoMVC. The former employs globally shared learnable weights to perform weighted fusion of different views, while the latter further introduces selective cross-view contrastive learning and uses global weights to regulate the alignment strength. However, both methods employ view-level fusion weights shared across all samples, primarily adjusting each view's contribution according to its overall importance, which makes it difficult to adaptively integrate evidence from different views for individual samples. In contrast, our method enables sample-wise adaptive cross-view interactions to construct consensus representations, allowing view representations to be jointly learned under the guidance of the final clustering objective. Meanwhile, an intra-view relational preservation constraint maintains the structural information of each view without imposing predefined consistency or complementarity relations across views.

2, We thank the reviewer for raising this question regarding the scope of our theoretical analysis. First, we clarify that the averaging operation in our analysis is not an arbitrarily chosen fusion strategy, but the optimal common center derived from a cross-view consistency constraint under the squared Euclidean distance. Therefore, the attenuation factor 1/V1/V corresponds to the specific case of equally weighted consistency constraints. More generally, the phenomenon we analyze is not limited to averaging itself, but stems from the compression of inter-view differences induced by distance-based consistency constraints. When the discriminative information of certain sample pairs is provided only by a subset of views, such constraints may weaken the corresponding view-specific discriminative information. Furthermore, if global view weights are introduced, the common center becomes a weighted average of view representations rather than a simple average, and the attenuation of view-specific discriminative information is consequently determined by the fixed view weights. However, different views may have different discriminative capabilities for different sample pairs, and globally shared weights cannot guarantee that the views carrying discriminative information for a given sample pair receive higher weights. This analysis also motivates our method design: discriminative information from different views should first be integrated at the sample level, followed by clustering structure learning on the resulting consensus representations. We have added a corresponding discussion to Appendix A.1 of the revised manuscript.

3, We thank the reviewer for this suggestion. Following the reviewer's recommendation, we conducted a controlled synthetic study with three settings featuring different cross-view information distributions. The results are reported in the table below. As the shared information across views decreases, CVCL and AWDCMVC, representing direct cross-view consistency constraints and global-to-local supervision, respectively, exhibit progressively declining performance. In contrast, DDMVC, which jointly learns consistency and diversity, shows an initial performance drop followed by relative stability, possibly benefiting from its diversity constraints that preserve view-specific information. Our method achieves the best clustering results across all three settings. These findings are consistent with our theoretical analysis, suggesting that enforcing cross-view consistency may hinder the utilization of complementary clustering information, whereas our method effectively integrates discriminative information across views. The experimental settings, results, and analysis have been added to Section 4.2 of the revised manuscript. In addition, upon acceptance of the paper, we will release the complete data generation code and the generated synthetic datasets in our public code repository to ensure reproducibility.

MethodFully SharedFully SharedFully SharedFully SharedPartially SharedPartially SharedPartially SharedPartially SharedDisjointDisjointDisjointDisjoint
accnmipurariaccnmipurariaccnmipurari
ours111111111111
AWDC-MVC11110.7030.7810.7030.5930.5460.6960.5690.443
DDMVC11110.7060.7600.7190.5690.6540.7770.6540.563
CVCL11110.5010.6070.5150.3510.2980.2610.3080.112

Finally, we hope that our clarifications, additional experiments, and corresponding revisions adequately address your concerns. We would appreciate your reconsideration of the overall assessment in light of these responses.

Reviewer_AhWoOpenReview ↗

Tanks for the results

The authors have conducted the experiments I was concerned about and provided a more detailed analysis, which helps address my previous concerns. I appreciate the effort put into the rebuttal and have decided to raise my score.

Since the paper aims to understand the underlying mechanisms, I believe that synthetic datasets can often provide clearer intuition than experiments on existing benchmarks, especially in graph-related research. Once the mechanism is clearly illustrated, showing that the findings also hold on standard benchmarks would make the overall argument more convincing. I encourage the authors to consider this approach when further improving the presentation of the paper.

Overall, I appreciate the authors' efforts to clarify and strengthen the work.

We appreciate the reviewer’s recognition of our work and the constructive suggestions provided throughout the review process.

We agree with the reviewer’s suggestion and have further reorganized the experimental section by moving the controlled synthetic experiments ahead of the comparisons on the main benchmark datasets. This organization provides a clearer progression from analyzing the underlying mechanisms to validating their implications in practical settings, establishing a more direct connection between the experimental results and the central research questions of the paper.

We thank the reviewer again for this valuable suggestion. It has helped us clarify the logical flow of the experimental section and improve the overall coherence and presentation of the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.