acceptodds
Withdrawn from ICLR 2027

Preserving Consistency and Progressively Integrating Diversity for Continual Multi-View Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual multi-view clustering (CMVC) updates clustering representations as new views arrive sequentially, requiring both the preservation of shared structures and the integration of complementary information. Existing methods typically preserve historical knowledge by constraining current representations to remain compatible with previous ones. Such compatibility constraints provide limited semantic guidance and often underexploit view-specific diversity. To address these limitations, we propose Preserving Consistency and Progressively Integrating Diversity for Continual Multi-View Clustering (PC-PID). PC-PID decomposes adjacent-view representations into a shared consistency component and view-specific diversity components. Contrastive knowledge transfer strengthens semantic dependence between current and historical consistency representations while suppressing redundant dependence between adjacent diversity representations, thereby preserving shared structures and separating view-specific information. Progressive tensor fusion constructs a local tensor from adjacent diversity representations and captures their high-order correlations as new views arrive. These two mechanisms are incorporated into a unified objective and optimized through an alternating algorithm. Experiments on eight benchmark datasets against 14 representative static and continual multi-view clustering methods demonstrate the effectiveness and efficiency of PC-PID.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Continual multi-view clustering, Consistency preservation, Progressive diversity integration, Contrastive knowledge transfer, Tensor fusion.

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_tTS4OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Preserving Consistency and Progressively Integrating Diversity for Continual Multi-View Clustering (PC-PID). The method decomposes adjacent-view representations into a shared consistency component and view-specific diversity components. Contrastive knowledge transfer uses HSIC to strengthen dependence between historical and current consistency representations while suppressing redundant dependence between diversity representations. Progressive tensor fusion organizes adjacent diversity representations into a local tensor and applies tensor nuclear norm regularization to capture high-order correlations. Experiments are conducted on eight multi-view datasets against 14 static and continual baselines. The paper claims effectiveness and efficiency for continual multi-view clustering.

Critical strengths

The paper addresses an important and challenging problem: continual multi-view clustering with sequentially arriving views.

The consistency-diversity decomposition provides a reasonable way to separate shared and view-specific information.

Contrastive knowledge transfer using HSIC is a principled alternative to simple distance-based transfer.

Progressive tensor fusion with a local tensor is an efficient way to capture high-order correlations without storing all historical views.

The method is evaluated on eight datasets and compared with many recent baselines.

Ablation studies, parameter sensitivity, and convergence analysis are provided.

Critical weaknesses

The hyperparameter selection protocol uses ground-truth labels. The paper states: ``The configuration with the highest mean ACC is selected.'' Since ACC requires ground-truth cluster labels, this constitutes information leakage and is not valid in an unsupervised clustering setting. The same issue applies to the ten-run mean-ACC selection criterion used for baselines. The authors should adopt an unsupervised model-selection criterion, such as silhouette score, modularity, or stability, and report results accordingly.

The HSIC definition appears unconventional and potentially incorrect. Definition~1 defines

HSIC(A,B)=1(m−1)2tr(KAKB),\mathrm{HSIC}(\mathbf{A},\mathbf{B}) = \frac{1}{(m-1)^2}\mathrm{tr}(\mathbf{K}_A\mathbf{K}_B),
where $\mathbf{K}_A = \mathbf{H}\mathbf{A}^\top\mathbf{A}\mathbf{H}$, $\mathbf{K}_B = \mathbf{H}\mathbf{B}^\top\mathbf{B}\mathbf{H}$, and $\mathbf{H} \in \mathbb{R}^{m \times m}$. Here $m$ is the latent dimension, not the number of samples. Standard HSIC operates on sample-level kernel matrices of size $n \times n$. Using $m \times m$ matrices measures dependence across latent dimensions, not across samples. This mismatch needs to be clarified, and the theoretical properties of the proposed HSIC-based transfer should be re-examined.

The theoretical analysis of contrastive knowledge transfer is incomplete. Proposition~1 only provides an upper bound and equality condition for HSIC, but does not establish that minimizing the proposed CKT loss actually transfers the desired common clustering structure. The connection between HSIC alignment and clustering performance is asserted rather than proved. A formal guarantee or at least a more rigorous justification is needed.

Table~1 is severely malformed and incomplete. Many method names and numerical values are concatenated (e.g., DAG-FC0.9330.873...''), several columns appear missing, and some results are unavailable (-''). It is impossible to verify the claimed superiority. For example, on Scene15, PC-PID achieves ACC 0.5510.551, while Orth-NTF achieves 0.8070.807, TLRLF4 achieves 0.8020.802, and TMC achieves 0.6230.623. On Caltech101-all, PC-PID achieves ACC 0.5510.551, while Orth-NTF achieves 0.5040.504, but TLRLF4 achieves 0.2770.277? The table is too corrupted to assess. Complete and correctly formatted tables with standard deviations and significance tests are required.

The claim of ``state-of-the-art performance'' is not supported by the presented results. On several datasets and metrics, PC-PID is outperformed by static methods that use the full view set. For instance, on Scene15 and Caltech101-all, PC-PID is worse than Orth-NTF and TLRLF4 in ACC. The paper should clearly distinguish between continual and static settings and avoid overclaiming.

The progressive tensor fusion is claimed to be efficient, but the complexity analysis omits the cost of tensor singular value thresholding (t-SVT). For a tensor of size n×m×2n \times m \times 2, t-SVT involves FFTs and matrix SVDs, whose complexity is not negligible. A detailed complexity breakdown and runtime comparison are needed.

The ablation study shows that removing CKT only decreases ACC by 0.0270.027, while removing PTF decreases ACC by 0.2770.277. This suggests that PTF is the dominant component and CKT contributes marginally. The paper should discuss whether CKT is necessary and whether the HSIC-based transfer is worth the added complexity.

The convergence analysis is not rigorous. The LOSS value temporarily increases around iterations 8--10, which is attributed to alternating block updates and adaptive penalty adjustment. However, no theoretical convergence guarantee is provided for the overall non-convex optimization. The stopping criterion and penalty parameter updates are heuristic. A formal convergence analysis or at least a more careful empirical study is needed.

The view-arrival order sensitivity experiment is only conducted on Scene15. The paper concludes that PC-PID is robust to order, but different orders lead to final ACC values ranging from 0.8680.868 to 0.9840.984. This is a large variation. The claim of robustness should be tempered, and more datasets should be tested.

Reproducibility is insufficient. The code is not provided in the submission, and the exact parameter configurations for baselines are not fully reported. The use of ACC for hyperparameter selection makes it impossible to reproduce the method in a truly unsupervised manner.

Questions for authors

How do you justify using ACC for hyperparameter selection in an unsupervised clustering task? This uses ground-truth labels and invalidates the unsupervised setting. Can you provide results using an unsupervised model-selection criterion?

Can you clarify the definition of HSIC? Standard HSIC uses sample-level kernel matrices of size n×nn \times n. Your definition uses m×mm \times m matrices, where mm is the latent dimension. What is the theoretical justification, and how does this measure sample dependence?

Can you provide a formal proof or stronger justification that minimizing the CKT loss transfers the common clustering structure? Proposition~1 only gives an upper bound and equality condition.

Can you provide complete, correctly formatted tables with all baseline results, standard deviations, and statistical significance tests? The current Table~1 is unreadable.

On Scene15 and Caltech101-all, PC-PID is outperformed by static methods such as Orth-NTF and TLRLF4. How do you reconcile this with the claim of state-of-the-art performance?

Can you provide a detailed complexity analysis including the cost of t-SVT? How does the runtime compare with baselines?

The ablation shows that removing CKT only reduces ACC by 0.0270.027, while removing PTF reduces it by 0.2770.277. Is CKT necessary? What is the marginal benefit of HSIC over simpler alternatives?

Can you provide a theoretical convergence guarantee for the overall optimization? The LOSS value increases temporarily, and no formal proof is given.

The view-arrival order experiment shows large variation in final ACC (0.8680.868 to 0.9840.984). How can users choose a good order in practice? Can you test more datasets?

Will the code and exact parameter configurations be released to ensure reproducibility?

Reviewer_7X2jOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PC-PID addresses continual multi-view clustering through consistency–diversity decomposition, HSIC-based knowledge transfer, and local low-rank tensor fusion. Experiments compare fourteen baselines on eight datasets, with additional ablation, efficiency, and view-order analyses.

Critical strengths
  1. The paper addresses the practical challenge of preserving historical knowledge while incorporating complementary information from incoming views.
  2. The evaluation includes component ablations and view-order analyses, supported by detailed optimization and experimental settings in the appendix.
Critical weaknesses
  1. Limited novelty in consistency–diversity modeling. CDIMVC already explores consistency and diversity in incremental multi-view clustering. The paper does not sufficiently establish a substantive advance beyond this framework through the addition of standard tensor regularization.
  2. The HSIC-based contribution is insufficiently distinguished. HSIC is also used in CDIMVC. The authors should clarify what is fundamentally different about their ratio-based transfer objective and why it provides stronger knowledge preservation than existing dependence-based formulations.
  3. Several references contain incorrect attribution. Reference [1] lists Songhe Feng instead of Yu Feng as first author; [2] lists Yawei Chen instead of Wei Zhang; and [3] lists Zisen Zhang instead of Jun Wang. The title of [3] also omits “Alignment.” These repeated errors require correction.
  4. Figure 1 does not clearly illustrate the claimed innovation. It depicts positive/negative relations and tensor stacking, but does not explain the distinctive cross-stage updates, retained historical information, or differences from existing CMVC pipelines.

[1] Yu Feng et al. Incremental Multi-View Clustering: Exploring Stream-View Correlations to Learn Consistency and Diversity. IEEE TKDE, 2025.

[2] Wei Zhang et al. Dual Anchor Graph Fuzzy Clustering for Multiview Data. IEEE Transactions on Fuzzy Systems, 2025.

[3] Jun Wang et al. Tensor Multi-Rank Constraint Guided Anchor-Wise Adaptive Alignment for Multi-View Clustering. IEEE TKDE, 2026.

Questions for authors
  1. When (\lambda_1=0), Eq. (35) sums the two residual representations, whereas the reconstruction objective requires their average. How is this update derived, and which version is implemented?
  2. Eqs. (17) and (19) imply (\mathrm{ACC}\leq\mathrm{Purity}). Why does Table 1 report (0.629>0.442) for CDIMVC on Caltech101-all? Is this a reporting or evaluation error?
Reviewer_pHJaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PC-PID for continual multi-view clustering. It reconstructs adjacent raw views using a shared consistency representation and separate diversity components. A ratio of kernel-dependence terms is intended to preserve historical consistency and suppress diversity redundancy; a two-slice tensor nuclear norm regularizes adjacent diversity representations. An alternating procedure produces a final representation combining the latest consistency with accumulated diversity. Evaluation spans eight datasets, 14 static/continual baselines, ablations, runtime comparisons, and Scene15 arrival-order analysis.

S1: The paper reconstructs adjacent views using a shared consistency component and separate diversity components. This design assigns explicit roles to common structure and view-specific information, providing a concrete framework for preserving historical knowledge while incorporating complementary information from incoming views. S2: The paper constructs a local tensor containing only the diversity representations of two adjacent views. This design keeps the tensor size independent of the number of previously received views, allowing local fusion without repeatedly constructing and processing a tensor over the complete view history. S3: The paper supplements benchmark comparisons with component ablations, runtime measurements, and all six view-arrival orders on Scene15. This setup examines which components contribute, how much computation is required, and how performance depends on view order, providing several complementary checks of the continual clustering design.

Critical strengths

S1: The paper reconstructs adjacent views using a shared consistency component and separate diversity components. This design assigns explicit roles to common structure and view-specific information, providing a concrete framework for preserving historical knowledge while incorporating complementary information from incoming views. S2: The paper constructs a local tensor containing only the diversity representations of two adjacent views. This design keeps the tensor size independent of the number of previously received views, allowing local fusion without repeatedly constructing and processing a tensor over the complete view history. S3: The paper supplements benchmark comparisons with component ablations, runtime measurements, and all six view-arrival orders on Scene15. This setup examines which components contribute, how much computation is required, and how performance depends on view order, providing several complementary checks of the continual clustering design.

Critical weaknesses

W1. The advertised solver is not a demonstrated optimizer of Eq. (13). Setting lambda_1 to zero, the consistency subproblem has minimizer C = (1/2) sum_i[X_i P_i^T - D_i] for the two adjacent views, whereas Eq. (35) returns the sum without the factor 1/2. With nonzero lambda_1, differentiating the ratio also involves the current denominator and diversity numerator; Eqs. (30), (34) and (35) instead use a historical-only normalization or omit these quantities. A fixed-point approximation could be legitimate, but its actual surrogate and connection to the stated objective must be supplied. Stabilization of the representation-change LOSS does not validate that connection.

W2. The empirical evidence contains contradictions requiring correction before the large gains can be trusted. Under the definitions in Eqs. (17) and (19), ACC cannot exceed Purity for the same predicted partition; nevertheless, Table 1 reports CDIMVC on Caltech101-all as ACC 0.629 and Purity 0.442, and ASIA as 0.271 and 0.129. Appendix C also selects dataset-specific configurations by the same benchmark's mean ACC, without describing separate validation. Together with absent uncertainty reporting and different stated search ranges, this leaves both the score records and the comparison protocol unresolved.

Questions for authors

Q1. Could the authors derive Eqs. (34)–(35) from Eq. (13), correcting the reconstruction-only factor and identifying every approximation to the dependence ratio? With both projected residuals equal to I and lambda_1 = 0, the exact C update is I, while the printed update is 2I. Please provide the actual implemented objective and its objective/residual traces after correction. Q2. In Eqs. (3)–(4), kernels are m-by-m and center latent dimensions. Independently permuting the sample rows of either input leaves A^T A or B^T B unchanged, despite changing sample-level cluster relations. How does this statistic preserve historical clustering geometry?
Q3. For the two-slice tensor in Eq. (10), Eq. (11) equals ||D_prev + D_cur||* + ||D_prev - D_cur||* under the stated FFT convention. What benefit is specific to progressive high-order fusion beyond this paired matrix regularizer?
Q4. The all-order Scene15 experiment is useful, but its final ACC ranges from 0.8680 to 0.9839. Could the authors report mean/dispersion over orders and compare retention against learning from the current view alone, including a weak or corrupted first view?

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

Related papers

Discussion (0)

Sign in to comment.