acceptodds
Under review as a conference paper at ICLR 2027

CORE: Coupled Representation and Cluster-Aware Topology for Scalable Multi-View Clustering

Abstract

As multi-view data continue to proliferate in real-world applications, anchor-based clustering provides an efficient paradigm for latent cluster discovery. Despite recent progress, existing methods still encounter the following limitations. First, they primarily emphasize multi-view consistency, whereas the learned shared structures lack sufficient cluster discriminability. Second, representation learning and topology optimization are commonly treated as separate processes, restricting mutual refinement during optimization. To address these issues, we propose Coupled Representation and Cluster-Aware Topology (CORE), a scalable multi-view clustering framework designed to leverage cross-view consistency to learn cluster-discriminative structures through unified optimization. Specifically, CORE jointly learns view-specific sample-anchor representations through robust reconstruction and imposes a low-rank tensor structure to capture high-order consistency across views. Meanwhile, an adaptive anchor topology is inferred from representations refined by cross-view consistency and regularized by spectral connectivity constraints to capture cluster-oriented structural information. Accordingly, representation learning and topology optimization are coupled to enable mutual refinement, yielding more discriminative structures for clustering. By performing structural learning in the compact anchor space, CORE avoids expensive sample-wise graph construction while maintaining favorable scalability on large-scale datasets. Extensive experiments on real-world benchmarks demonstrate the effectiveness of CORE across diverse clustering scenarios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view Learning, Clustering, Topology Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_76kyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

CORE is a scalable multi-view clustering method that jointly learns sample-anchor representations and an anchor-level topology. Cross-view information is captured by applying tensor regularization to the view-specific representations, while spectral connectivity is imposed on the anchor graph to encourage cluster-oriented structure. The optimization is performed in the compact anchor space, which avoids constructing full sample-wise graphs. Experiments on ten multi-view datasets show strong clustering results and competitive computational efficiency.

Critical strengths
  • The use of an anchor-level topology provides an effective way to incorporate explicit structural information without constructing expensive sample-wise graphs. This design preserves the advantages of graph-based clustering while keeping the computational cost manageable for larger datasets.

  • The spectral connectivity constraint gives the learned anchor graph a clear clustering-oriented objective. Rather than relying only on cross-view consistency, CORE explicitly encourages the topology to reflect latent cluster structure, which is a meaningful addition to conventional anchor-based multi-view clustering.

  • CORE achieves strong performance across datasets with substantially different sample sizes and numbers of clusters, while maintaining competitive running time. The consistency of these results suggests that the proposed framework provides a favorable balance between clustering effectiveness and computational efficiency.

Critical weaknesses
  • The learned anchor topology is mainly evaluated through qualitative visualization, while its structural quality is not quantitatively analyzed. Since spectral connectivity is a key component of CORE, additional evidence on the resulting graph structure would further support the effectiveness of the topology learning strategy.

  • The number of anchors mm is selected from a predefined candidate set, but its influence on clustering performance and computational efficiency is not explicitly investigated. A more detailed analysis of the trade-off associated with different anchor numbers would improve the practical understanding of CORE.

Questions for authors
  • Can the authors provide quantitative analysis of the learned anchor topology, such as the number of connected components or the Laplacian spectrum, to complement the current visualization results?

  • How sensitive is CORE to the number of anchors mm, and how does mm affect the trade-off between clustering performance and computational efficiency?

  • The spectral embedding FF is learned to enforce the cluster-aware structure of the anchor topology, while the final clustering is performed on the concatenated sample-anchor representations. Can the authors further clarify the role of FF?

  • Can the authors provide further discussion on how the anchor number is selected in practice and whether a similar range of mm works consistently across datasets with different scales and numbers of clusters?

3 comments

We sincerely thank the reviewer for the careful evaluation and valuable feedback. The questions regarding topology validation and anchor selection are particularly helpful, and we have conducted additional experiments to address them.

W1 & Q1: Quantitative Validation of the Learned Topology.

We agree that visualization alone is insufficient to fully demonstrate the structural quality of the learned anchor topology. Following the reviewer's suggestion, we conducted an additional quantitative analysis of the Laplacian spectrum and connected components, comparing CORE with a variant that removes spectral connectivity while retaining adaptive topology learning. The results are summarized below:

DatasetMethodComponents∑i=1cλi\sum_{i=1}^{c}\lambda_iλc+1\lambda_{c+1}Eigengap
NUS (c=12c=12)CORE120.00001.00981.0098
w/o Spectral111.07161.00770.00003
ALOI100 (c=100c=100)CORE1000.00001.00171.0017
w/o Spectral194.05601.00640.00000

The results demonstrate that CORE achieves the desired number of connected components with near-zero smallest cc eigenvalues and a clear eigengap, whereas removing spectral connectivity produces a single connected component. This confirms that spectral regularization effectively guides the learned topology toward the expected cluster-oriented structure, providing quantitative evidence beyond the existing visualizations. We will include this analysis in the final version.

W2 & Q2: Anchor-Number Sensitivity and Performance–Efficiency Trade-off.

We thank the reviewer for the valuable suggestion regarding anchor-number sensitivity. To investigate the influence of mm on clustering performance and computational efficiency, we conducted additional experiments on BDGP and NUS with m/c∈{2,3,5,10}m/c\in\{2,3,5,10\}, while keeping the remaining settings unchanged. The results are reported below:

DatasetMetric2c2c3c3c5c5c10c10c
BDGPACC (%)98.2898.2898.2098.20
NMI (%)94.9394.9194.7094.69
Time (s)4.14.24.35.0
NUSACC (%)87.6592.8395.3897.20
NMI (%)77.9987.0791.7394.69
Time (s)3.53.85.08.3

The results reveal different sensitivity patterns across datasets. On BDGP, a relatively small anchor set is sufficient to characterize the underlying structure, resulting in stable clustering performance as m increases. In contrast, NUS benefits from additional anchors, which provide greater capacity to capture complex structural relationships. Meanwhile, the running time increases with mm on both datasets. These observations highlight the importance of balancing structural representation capacity and computational efficiency. We will include these additional results and further discuss this trade-off in the final version.

Q3: Role of the Spectral Embedding FF.

The spectral embedding F∈Rm×cF\in\mathbb{R}^{m\times c} serves as an anchor-level structural variable rather than the final sample-level clustering representation. Specifically, minimizing Tr(F⊤LGF)\mathrm{Tr}(F^\top L_GF) under F⊤F=IcF^\top F=I_c selects the eigenvectors corresponding to the smallest eigenvalues of LGL_G, characterizing the connectivity structure of the anchor graph. The pairwise distances between the rows of FF further participate in updating GG, encouraging cluster-oriented relationships among anchors.

The structural information encoded by FF is subsequently transferred to BvB^v through Laplacian regularization. Since the final clustering requires sample-level assignments, the refined representations are concatenated as H=[B1,…,BV]H=[B^1,\ldots,B^V], followed by SVD and kk-means. Consequently, FF provides cluster-aware structural guidance during optimization, while HH integrates the refined multi-view representations for final clustering. We will clarify their distinct roles in the revised manuscript.

Q4: Practical Selection of Anchor Numbers.

In practice, we recommend m∈{2c,3c,5c,10c}m\in\{2c,3c,5c,10c\} as a candidate range, with the selection guided by clustering stability, graph spectral properties, and computational cost. As illustrated by the sensitivity results on BDGP and NUS, different datasets may favor different anchor ratios. Furthermore, ensemble clustering across multiple anchor settings could potentially integrate structural information at different granularities and reduce dependence on a single anchor number. We will further investigate this direction in future work.

Overall, we appreciate the reviewer's constructive feedback and hope our responses reinforce your favorable assessment of the paper.

The added explanations and experiments address my concerns. I therefore keep my score.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer 76ky,

Thank you again for your careful review and constructive suggestions. We appreciate your acknowledgment that our additional explanations and experiments have addressed your concerns regarding topology validation, anchor-number sensitivity, and spectral embedding.

We will incorporate these additional results and clarifications into the revised manuscript. We also appreciate your decision to maintain a positive assessment of our work.

Thank you very much for your time and consideration.

Best regards,

The Authors

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

To achieve efficient and effective multi-view clustering, this paper develops CORE, a unified framework that jointly explores cross-view representation consistency and cluster-oriented anchor structures. At the representation level, CORE employs robust reconstruction to obtain sample-anchor relationships for individual views, while a non-convex low-rank tensor constraint is introduced to characterize their high-order dependencies across multiple views. At the structural level, the resulting representations guide the adaptive construction of an anchor topology, whose spectral connectivity is explicitly regulated to reveal the underlying cluster structure. Rather than treating these two processes independently, CORE integrates them into a joint optimization scheme, allowing the learned representations and anchor topology to progressively improve each other. Evaluations across ten benchmark datasets demonstrate that CORE achieves competitive clustering accuracy with favorable computational efficiency. Additional ablation experiments substantiate the effectiveness of its individual components.

Critical strengths
  • This paper presents a coherent coupling between sample-anchor representation learning and cluster-aware anchor topology optimization. Importantly, the learned topology is not treated as an auxiliary structure but directly participates in refining the representations used for clustering.

  • The integration of robust reconstruction with non-convex low-rank tensor regularization provides a well-motivated mechanism for suppressing corruption while capturing high-order cross-view consistency, thereby enabling more effective exploitation of information across multiple views.

  • The experimental evaluation is comprehensive. CORE is assessed on datasets with varying scales and numbers of clusters, with the main comparisons further supported by ablation, runtime, convergence, parameter sensitivity, and visualization studies.

Critical weaknesses
  • Although CORE integrates sample-anchor representation learning, tensor regularization, topology learning, and spectral connectivity within a unified framework, the paper could more clearly articulate what additional benefit is brought by coupling these components beyond their individual contributions.

  • The paper could further clarify the complementary roles of tensor regularization and spectral connectivity. While tensor regularization is intended to capture high-order cross-view consistency and spectral connectivity introduces cluster-oriented structural information, the interaction between these two mechanisms and how they jointly improve the learned representations could be explained more explicitly.

Questions for authors
  • Can the authors further clarify what additional benefit is obtained by coupling sample-anchor representation learning with anchor topology optimization, beyond the individual contributions of the two components?

  • Can the authors provide more intuition on how tensor regularization and spectral connectivity play complementary roles in improving the learned representations?

  • Can the authors further explain how the learned anchor topology feeds back into the sample-anchor representations during alternating optimization and contributes to improved clustering discriminability?

2 comments
AuthorsOpenReview ↗

Clarifications on Representation–Topology Coupling and Complementary Structural Learning

We appreciate the reviewer's positive assessment and constructive feedback. The following responses further clarify the key contributions of CORE and provide additional experimental evidence.

W1 & Q1: Benefits of Representation–Topology Coupling.

We thank the reviewer for the insightful question regarding the contribution of the coupling mechanism. The key contribution of CORE lies in enabling mutual refinement between sample-anchor representations and cluster-aware topology, rather than simply combining individual regularization terms. Specifically, the tensor-refined representations BvB^v are used to adaptively construct anchor relationships based on their response patterns across multiple views, while spectral connectivity further guides the learned topology toward cluster-oriented structures. Importantly, the resulting graph Laplacian LGL_G directly regularizes subsequent updates of BvB^v, encouraging strongly connected anchors to exhibit coherent sample-response patterns. The refined representations are then used to update the topology, establishing a bidirectional optimization process that progressively incorporates cluster-oriented structural information into representation learning.

To further evaluate the effectiveness of topology feedback, we conducted an additional experiment by removing the topology-dependent contribution from the BvB^v update. The ACC results (%) are reported below:

DatasetCORENoFeedback
NUS97.2029.20
ALOI10087.2279.05

The results show that CORE substantially outperforms NoFeedback, indicating that learning the anchor topology alone is insufficient to achieve the same improvements. Instead, incorporating the learned cluster-oriented relationships into representation refinement is essential for improving clustering discriminability in these experiments. We will further clarify the coupling mechanism in the final version.

W2 & Q2: Complementarity of Tensor Regularization and Spectral Connectivity.

We appreciate the reviewer's suggestion to clarify the interaction between these two components. In CORE, the non-convex low-rank tensor regularization captures high-order correlations among BvB^v, preserving shared structural information and reducing view-specific inconsistencies. Since the anchor topology GG is constructed from the pairwise distances between anchor-response columns, these refined representations provide more reliable structural information for topology learning. Subsequently, the spectral connectivity constraint guides GG toward cc connected components, introducing explicit cluster-oriented relationships. Through the structure-coupling term ∑vTr(BvLGBv⊤)\sum_v\mathrm{Tr}(B^vL_GB^{v\top}), these relationships further guide representation refinement, thereby integrating cross-view consistency with clustering discriminability.

To further examine their contributions, we conducted additional experiments. Specifically, Neither removes both tensor regularization and spectral connectivity; Tensor Only and Spectral Only retain only the corresponding component. The ACC results (%) are presented below:

DatasetNeitherTensor OnlySpectral OnlyCORE
NUS24.0596.3224.2097.20
ALOI10069.1986.6366.8887.22

The results show that tensor regularization helps learn consistent sample-anchor representations across views, providing reliable anchor relationships for topology construction. In contrast, spectral connectivity alone can constrain the graph structure but cannot ensure that the learned anchor relationships accurately reflect the underlying clusters. By combining both mechanisms, CORE further organizes these relationships into cluster-oriented structures, leading to improved clustering performance. We will further clarify the coupling mechanism in the final version.

Q3: Topology Feedback for Representation Refinement.

The learned anchor topology directly influences representation learning through the update of BvB^v. Specifically, the coupling term γTr(BvLGBv⊤)\gamma\mathrm{Tr}(B^vL_GB^{v\top}) contributes 2γBvLG2\gamma B^vL_G to the gradient, yielding the update involving (μAA⊤+ξIm+2γLG)−1(\mu AA^\top+\xi I_m+2\gamma L_G)^{-1}. This term penalizes differences between strongly connected anchor responses, encouraging connected anchors to maintain coherent sample-response patterns. Since spectral connectivity guides the topology toward cluster-oriented structures, the regularization promotes within-cluster consistency while limiting smoothing across weakly connected anchor groups. Consequently, cluster-aware structural information is incorporated into BvB^v to enhance representation discriminability. We will further clarify this mechanism in the revised manuscript.

We hope these clarifications and additional experimental results address your concerns, and sincerely hope you will maintain your positive assessment of our work.

AuthorsOpenReview ↗

Follow-up on Additional Experiments and Clarifications

Dear Reviewer uJrP,

Thank you again for your positive assessment and constructive feedback. As the discussion period approaches its end, we would like to briefly summarize our additional clarifications and experiments.

To address your concerns regarding representation–topology coupling, we conducted a NoFeedback ablation to demonstrate the importance of topology feedback. We also provided additional ablation results to clarify the complementary roles of tensor regularization and spectral connectivity, together with a detailed explanation of how the learned topology guides representation refinement.

We hope these additional results and explanations have adequately addressed your concerns and will support a positive final assessment of CORE. If any questions remain, we would be glad to provide further clarification.

Thank you for your time and consideration.

Best regards,

The Authors

Reviewer_DNK5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an anchor-based framework for scalable multi-view clustering. The model factorizes each view using a shared latent anchor basis together with view-specific sample-anchor representations and projections. Cross-view consistency is further captured through non-convex low-rank tensor regularization, while a shared anchor topology is adaptively learned and regularized by spectral connectivity to promote cluster-oriented structure. The resulting representations are used for final clustering after dimensionality reduction. Experiments on ten multi-view benchmarks show strong performance and competitive efficiency.

Critical strengths

The factorized representation model provides a clear separation between shared latent anchor structure and view-specific transformations, offering a compact way to accommodate heterogeneous feature spaces while retaining a common representation basis.

Learning structural relationships in the anchor space avoids the cost of constructing sample-wise graphs, while the spectral connectivity term introduces an explicit cluster-oriented inductive bias into the learned topology.

The method is supported by consistently promising results across a diverse collection of multi-view datasets. The optimization procedure is also described in detail, making the overall framework relatively easy to follow and reproduce.

Critical weaknesses

Different views contribute to the topology regularization in a largely uniform manner, without an explicit mechanism for accounting for differences in view quality or reliability. This may make the learned consensus structure sensitive to noisy or weakly informative views when the quality of individual views varies substantially.

The sample-anchor representation BvB^v is learned without non-negativity or simplex constraints. While this provides additional modeling flexibility, the paper does not discuss how the unconstrained representation affects the interpretation of sample-anchor relationships or whether conventional anchor-graph constraints would lead to different behavior.

Questions for authors
  • What motivates the use of a single shared anchor topology GG across all views, rather than learning view-specific topologies and subsequently extracting their consensus structure?

  • How robust is CORE when one or more views are substantially noisier or less informative than the others? Can the authors provide a controlled noisy-view experiment to examine whether the shared topology is dominated by unreliable views?

  • Since BvB^v is interpreted as a sample-anchor representation, what is the motivation for leaving it unconstrained? Would imposing non-negativity or simplex constraints significantly affect clustering performance?

  • CORE assumes a shared anchor basis AA across all views while using RvR^v to accommodate view-specific feature spaces. Can the authors provide further intuition on why this factorization is sufficiently flexible for highly heterogeneous views?

5 comments
AuthorsOpenReview ↗

Clarifications on Shared Topology, View Reliability, and Anchor Representation Learning

We sincerely thank the reviewer for the positive assessment and constructive comments. We provide the following clarifications and additional experimental results.

W1 & Q1–Q2: Shared Anchor Topology and View Reliability.

Q1: We thank the reviewer for the thoughtful consideration of structural relationships across different views. The shared topology GG is designed to capture consistent structural information across multiple views and reveal the underlying cluster structure. Specifically, view-specific complementarity is preserved through BvB^v, RvR^v, and EvE^v, while cluster-oriented consistency is enhanced by imposing the spectral relaxation of the rank constraint rank(LG)=m−c\mathrm{rank}(L_G)=m-c on the shared topology. Consequently, learning a unified cluster-aware topology allows structural information from different views to be jointly exploited, providing more discriminative guidance for representation learning than treating their topologies independently.

Q2: We thank the reviewer for the valuable suggestion regarding view reliability. To further evaluate CORE's robustness to unreliable views, we conducted additional controlled experiments. Specifically, Gaussian noise was added to the first view, scaled by the standard deviation of each feature, while the remaining views were kept unchanged. The ACC results (%) under different noise levels are reported below:

DatasetClean0.10.20.30.51.0
BDGP98.2898.2498.2498.2498.0898.08
NUS97.2097.0396.9096.7596.8497.06
Animal91.5191.0890.9290.7592.3291.10

The results demonstrate that CORE maintains stable clustering performance across different noise levels. Accordingly, by integrating robust reconstruction with low-rank tensor regularization, CORE mitigates the influence of corrupted observations while preserving reliable cross-view structures, thereby reducing the impact of view-specific noise on clustering.

W2 & Q3: Unconstrained Sample-Anchor Representation.

We thank the reviewer for the constructive suggestion regarding representation constraints. In CORE, BvB^v is learned through the factorization Xv=BvARv+EvX^v=B^vAR^v+E^v, where AA denotes the shared latent anchor basis and RvR^v maps it into the view-specific feature space. Therefore, BvB^v represents latent sample-anchor response coefficients rather than probabilistic affinities, allowing signed values to flexibly capture latent structural relationships.

Following the reviewer's suggestion, we conducted additional experiments to examine the effect of imposing non-negativity on BvB^v. The ACC results (%) are reported below:

DatasetCORENonnegative
BDGP98.2898.20
NUS97.2096.39
ALOI10087.2287.16

As shown in the table, CORE achieves better clustering performance across all three datasets. By allowing signed response coefficients, CORE can capture richer latent structural information, thereby enhancing representation discriminability for clustering.

Q4: Flexibility of the Shared Anchor Basis.

We thank the reviewer for the insightful question regarding view heterogeneity. In CORE, the shared anchor basis AA is introduced to capture common latent anchor semantics through the factorization Xv=BvARv+EvX^v=B^vAR^v+E^v. Specifically, AA establishes a unified latent anchor space, while the view-specific projections RvR^v map this shared basis into different feature spaces, accommodating variations in feature dimensions and distributions. Meanwhile, BvB^v preserves view-dependent sample-anchor relationships, and EvE^v accounts for reconstruction residuals. Importantly, sharing AA establishes implicit anchor correspondences across views, reducing the need for additional cross-view anchor alignment procedures and avoiding inconsistencies arising from independently learned anchor bases. Consequently, this factorization preserves view-specific information while facilitating consistent structural learning across heterogeneous views.

We hope these clarifications and additional experimental results address your concerns, and sincerely hope you will maintain your positive assessment of our work.

Reviewer_DNK5OpenReview ↗

W1 & Q1–Q2: Shared Anchor Topology and View Reliability

In clustering task, estimating the rank constraint cc poses a significant challenge. The authors might discuss some practical methods to assist with this estimation, rather than simply specifying the number of clusters cc. For instance, could the performance on downstream tasks be used to inform the estimation of the rank constraint, or could a self-supervised approach be designed for this purpose?

AuthorsOpenReview ↗

Automatic clustering number determination

We sincerely thank the reviewer for the constructive suggestion regarding automatic clustering number determination. Specifically, CORE currently follows the standard clustering setting where the number of clusters cc is predefined, leading to the constraint rank(LG)=m−c\mathrm{rank}(L_G)=m-c. For scenarios where cc is unknown, we consider two potential directions worth exploring in future work:

Graph-Based Cluster Number Estimation. For unknown cc, a preliminary anchor graph can first be learned from the coefficient matrices Bvv=1V{\mathbf{B}^v}_{v=1}^V without imposing spectral connectivity constraints:​

G(0)=arg⁡min⁡γ∑v=1VTr⁡(BvLGBv⊤)+β∣∣G∣∣F2\mathbf{G}^{(0)}=\arg\min\gamma\sum_{v=1}^{V}\operatorname{Tr}(\mathbf{B}^v\mathbf{L}_G\mathbf{B}^{v\top})+\beta||\mathbf{G}||_F^2

Given the ascending eigenvalues λ1≤⋯≤λm\lambda_1\leq\cdots\leq\lambda_m of the resulting graph Laplacian, the cluster number could be estimated using the eigengap criterion:​ c^=arg⁡max⁡1≤k≤Kmax⁡(λk+1−λk)\hat{c}=\arg\max_{1\leq k\leq K_{\max}}(\lambda_{k+1}-\lambda_k). The estimated c^\hat{c} could then be incorporated into the spectral connectivity constraint rank(LG)=m−c^\mathrm{rank}(\mathbf{L}_G)=m-\hat{c} for subsequent joint optimization.

Self-Supervised Cluster Number Selection. Another possible strategy is to exploit cross-view consistency and clustering stability as self-supervised signals. Specifically, different candidate cluster numbers can be evaluated by measuring the agreement of clustering assignments across views and under data perturbations. A suitable cc could then be selected by considering assignment consistency and cluster balance. This would enable the rank constraint to be determined through the learned representations rather than a predefined cluster number.

We hope our response provides some potential directions for addressing this concern. We will further discuss these extensions in the Future Work section of the revised manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the Response (1/2)

We sincerely thank all reviewers and the Area Chair for their careful evaluations and constructive feedback. We have provided detailed responses in the respective review threads and conducted additional experiments to address the main concerns. The key findings and clarifications are summarized below.

Benefits of Representation–Topology Coupling: To address the reviewer's concern regarding the additional benefits of coupling representation learning with topology optimization, we conducted a NoFeedback ablation, which retains anchor topology learning but removes its influence on representation updates. The ACC decreases from 97.20% to 29.20% on NUS and from 87.22% to 79.05% on ALOI100. These results demonstrate that learning the topology alone is insufficient, while incorporating the learned structural information into representation refinement is essential for improving clustering performance.

Complementarity of Tensor Regularization and Spectral Connectivity: To clarify how these two mechanisms complement each other, we conducted an additional ablation study comparing their individual and combined contributions. On NUS and ALOI100, Tensor Only achieves ACC values of 96.32% and 86.63%, respectively, while incorporating spectral connectivity further improves them to 97.20% and 87.22%. These results indicate that tensor regularization provides reliable cross-view representations, while spectral connectivity further refines their cluster-oriented structural relationships, allowing both mechanisms to jointly improve clustering performance.

Robustness to Unreliable Views: To address concerns about whether noisy views could negatively affect the shared topology, we conducted additional experiments by introducing different levels of Gaussian noise into one view while keeping the remaining views unchanged. Even at the highest noise level, CORE maintains ACC values of 98.08% on BDGP, 97.06% on NUS, and 91.10% on Animal, remaining close to their clean-data performance. These results demonstrate that CORE maintains stable clustering performance under view-specific corruption, supporting the effectiveness of its robust representation learning mechanism.

Unconstrained Representations and View Heterogeneity: To clarify the motivation for unconstrained sample-anchor representations, we explained that BvB^v represents latent response coefficients rather than probabilistic anchor affinities. We further conducted nonnegativity-constrained experiments, showing that the original CORE achieves better ACC. Regarding view heterogeneity, we clarified that the shared basis AA establishes common latent anchor semantics, while view-specific projections RvR^v accommodate different feature spaces. These findings and explanations support the flexibility of signed representations and the shared factorization in modeling heterogeneous multi-view data.

Quantitative Validation of the Learned Topology: To address the concern that topology quality was mainly supported by visualizations, we conducted additional analyses of connected components and Laplacian spectra, comparing CORE with a variant without spectral connectivity. CORE achieves 12 connected components on NUS and 100 on ALOI100, matching their respective cluster numbers, with near-zero smallest cc eigenvalues and clear eigengaps. In contrast, removing spectral connectivity results in a single connected component on both datasets. These results quantitatively confirm that spectral connectivity effectively guides the learned graph toward the intended cluster-oriented structure.

AuthorsOpenReview ↗

Summary of the Response (2/2)

Anchor-Number Sensitivity and Computational Efficiency: To investigate how anchor numbers affect clustering performance and computational cost, we conducted additional experiments with m/c∈{2,3,5,10}m/c\in\{2,3,5,10\}. On BDGP, ACC remains stable between 98.20% and 98.28%, while on NUS, ACC increases from 87.65% to 97.20% as the anchor number grows. Meanwhile, running time increases on both datasets. These results demonstrate that different datasets require different levels of anchor representation capacity, highlighting the importance of balancing clustering effectiveness and computational efficiency. We also discussed practical anchor selection and potential ensemble extensions.

Role of Spectral Embedding and Automatic Cluster-Number Determination: To clarify why the spectral embedding FF is learned separately from the final clustering representations, we explained that FF guides the connectivity structure of the anchor topology, while the refined sample-anchor representations are used to obtain sample-level cluster assignments. In response to the additional concern regarding unknown cluster numbers, we further proposed two potential technical directions: graph-based estimation using Laplacian eigengaps and self-supervised selection based on cross-view consistency and clustering stability. These provide feasible directions for extending CORE beyond the predefined cluster-number setting.

The additional experimental results and methodological clarifications will be incorporated into the revised manuscript. We hope these efforts will receive favorable consideration and support from the reviewers and Area Chair.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.