acceptodds
Under review as a conference paper at ICLR 2027

Multi-Granularity Relation Transfer with Shared Anchors for Incomplete Multi-View Clustering

Abstract

Multi-granularity modeling aims to capture complementary cluster structures from fine to coarse resolutions to enhance incomplete multi-view clustering. Unfortunately, view incompleteness disrupts such modeling by inducing two major problems: cross-view reference inconsistency and cross-granularity structural instability. The first is reference inconsistency, where missing views make relational structures depend on different observed sample subsets, making structures from different views and granularities difficult to integrate. Considering this, we introduce shared anchors as a unified relational reference and construct masked sample-anchor assignments, which are further aggregated into a sample-complete global reference without imputing missing views. The second is multi-granularity instability: fine-grained structures are sensitive to missingness and noise, while coarse-grained structures may suffer from semantic drift. Therefore, we develop an anchor-induced multi-granularity relation transfer mechanism to extract fine-to-coarse structures on compact anchor graphs and transfer them to samples. By integrating sparse local geometry, our model preserves supported neighborhood evidence and promotes local consistency. Additionally, adaptive low-rank consensus learning adjusts the relative view-granularity contributions according to their consistency with the evolving consensus. Extensive experiments conducted on multiple benchmark incomplete multi-view datasets demonstrate the superiority of the proposed method over existing state-of-the-art approaches.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Shared Anchors, Multi-Granularity Relation Transfer, Adaptive Low-Rank Consensus Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_n5Q6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AMGC is a graph-based method for incomplete multi-view clustering that leverages anchor partitions at multiple resolutions to infer sample relationships, refines local structure using mutual k-nearest-neighbor graphs, and directly learns a low-rank consensus embedding with adaptive weighting. Extensive experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper addresses clustering with missing views and uses anchor graphs to reduce the computational cost.
  2. The paper conducts experiments on five benchmark datasets, including comparisons and ablation studies, to evaluate the effectiveness of the proposed method.
Critical weaknesses
  1. The methodological novelty appears limited. Shared anchors or shared anchor graphs for IMVC already appear in DVSAI [1] and FIML [2]. VMAL [3] transfers anchor cluster information to samples, and MGE [4] combines partitions at multiple resolutions. The additional contribution needs more justification.

[1] Yu, S., et al. DVSAI: Diverse View-Shared Anchors Based Incomplete Multi-View Clustering. AAAI, vol. 38, no. 15, pp. 16568–16577, 2024.

[2] Qin, Y., Feng, G., and Zhang, X. Fast Incomplete Multi-view Clustering by Flexible Anchor Learning. ICML, PMLR, vol. 267, pp. 50235–50245, 2025.

[3] Wei, X., et al. View-Adaptive Multi-Granularity Anchor Learning for Multi-View Clustering. IEEE TIP, vol. 35, pp. 3113–3126, 2026.

[4] Yang, J., et al. Multi-view Clustering via Multi-granularity Ensemble. IJCAI, pp. 6794–6802, 2025.

  1. Many anchor-graph-based methods have been proposed for incomplete multi-view clustering, but the baselines give limited coverage to this line of work.

  2. All five datasets contain 10,000 samples or fewer, which is insufficient to establish the claimed scalability. Evaluation on a dataset with around 100,000 samples, such as YoutubeFace_Sel used by DVSAI [1], would provide a more convincing test.

  3. The anchor selection procedure favors samples observed in more views. When some classes have more missing views than others, could these classes be underrepresented among the selected anchors? Please evaluate anchor class coverage when few fully observed samples are available.

  4. Are the selected anchor samples guaranteed to be observed in every view? If not, an anchor missing from a view has a zero column in that view’s assignment matrix and contributes no information to Eq. (5). How does this affect clustering performance?.

  5. The manuscript gives dataset-specific settings without explaining how AMGC or baseline parameters were selected. Table 1 also omits standard deviations.

Questions for authors
  1. What is the main methodological advance over existing methods?

  2. Could the authors compare AMGC with additional anchor-based IMVC methods and evaluate it on larger-scale datasets to further assess its scalability, particularly in terms of runtime and memory consumption?

  3. Are the selected anchors guaranteed to be observed in every view? If not, how does missing anchor information affect clustering performance? Please also evaluate anchor class coverage when few fully observed samples are available.

  4. How were the parameters of AMGC and the baselines selected? Were ground-truth labels used for parameter tuning?

3 comments
AuthorsOpenReview ↗

A Gentle Follow-up (Submission #12155)

Dear Reviewer n5Q6,

With the discussion period nearing its conclusion, we would like to express our sincere appreciation for your insightful suggestions and careful evaluation. We hope our additional experiments and clarifications have sufficiently addressed the concerns you raised. If there are any remaining questions, please do not hesitate to let us know. We would be glad to provide further explanations. If our responses have resolved your concerns, we would sincerely appreciate your consideration of updating your rating.

Thank you again for your time and consideration.

Best regards,

Authors

I appreciate the authors' efforts to address my concerns throughout the discussion. Although some concerns remain, I am willing to raise my score.

AuthorsOpenReview ↗

Thank You for Your Consideration!

We sincerely thank the reviewer for your thoughtful feedback and constructive engagement throughout the discussion! We truly appreciate your willingness to raise the score and your valuable support! Thank you again for your time and consideration!

Reviewer_ss4qOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Anchor-induced Multi-Granularity Consensus (AMGC), a framework for incomplete multi-view clustering that addresses cross-view reference inconsistency and structural instability across different granularities.

The proposed method consists of three main components. First, a shared-anchor reference construction mechanism selects representative anchors based on view completeness and feature diversity, then aggregates masked sample-anchor assignments to establish a unified relational reference without imputing missing views. Second, an anchor-induced multi-granularity relation transfer mechanism performs spectral partitioning at different resolutions in the anchor space and transfers the resulting structures to sample-level relations. Local geometric information is further incorporated through mutual k-nearest-neighbor graphs to preserve neighborhood consistency. Third, an adaptive low-rank consensus learning strategy integrates view-specific and global multi-granularity relations, dynamically adjusting their contributions according to fitting consistency while enabling efficient matrix-free optimization.

The authors provide theoretical analyses of relation-transfer stability, spectral perturbation, and optimization properties. Experiments on five benchmark datasets under four missing-view rates (0.1–0.7) demonstrate consistently strong performance against nine competing methods using ACC, NMI, and ARI. Ablation studies, hyperparameter analyses, and computational efficiency comparisons further support the effectiveness and practical applicability of the proposed framework.

Critical strengths
  1. Strong and Consistent Empirical Performance. AMGC demonstrates consistently competitive clustering performance across five benchmark datasets, four missing-view rates, and three evaluation metrics (ACC, NMI, and ARI), outperforming nine representative baselines in the reported results. The ablation studies further demonstrate the contributions of the proposed components. These results provide substantial empirical evidence of the method's effectiveness under the evaluated incomplete multi-view settings.

  2. Coherent Framework with Theoretical and Computational Support. The paper presents a well-structured framework that integrates shared-anchor reference construction, multi-granularity relation transfer, and adaptive low-rank consensus learning to address incomplete multi-view clustering. The theoretical analysis provides insights into relation-transfer stability and optimization properties, while the matrix-free formulation enables efficient consensus optimization. Together with the reported runtime and memory comparisons, these aspects strengthen the technical credibility and practical relevance of the proposed approach.

Critical weaknesses
  1. Insufficient Validation Under Challenging Missing-View Settings. The experimental protocol in Appendix B.2 preserves a substantial proportion of fully observed samples, which may favor the proposed completeness-aware shared-anchor selection strategy. Although Appendix C.5 includes an additional class-dependent missingness experiment, the robustness of the full framework under limited cross-view overlap or the absence of fully observed samples remains insufficiently evaluated. This limits the extent to which the reported results substantiate the method's effectiveness in more challenging incomplete multi-view scenarios.
Questions for authors

Q1. Novelty of Multi-Granularity Relation Transfer. The proposed multi-granularity relation transfer mechanism shares similarities with existing multi-granularity clustering and anchor-based methods, such as MGE (IJCAI 2025) and VMAL (IEEE TIP 2026). Could the authors clarify the key methodological differences and explain what additional capabilities the proposed mechanism provides beyond existing anchor-to-sample structure transfer and multi-granularity aggregation approaches?

Q2. Robustness Under More Challenging Missing-View Settings. According to Appendix B.2, the adopted missing-view protocol preserves a substantial proportion of fully observed samples, even at high missing rates. Given that shared-anchor construction prioritizes samples with higher view completeness, how would AMGC perform when fully observed samples are scarce or unavailable, or when cross-view sample overlap is limited? Could the authors provide additional evidence under such settings to substantiate the robustness of the proposed shared-anchor reference mechanism?

Q3. Reliability of Adaptive Granularity Weighting. Appendix A.3 acknowledges that inverse-error weighting may favor relations that are easier to approximate under the rank-KK constraint rather than those that are semantically more reliable. Although Appendix C.4 demonstrates improvements over uniform weighting, could the authors further clarify how the learned weights reflect the actual quality of different granularity-specific structures? In particular, to what extent might the rank-dependent fitting difficulty influence the learned weights and the resulting clustering performance?

1 comment
AuthorsOpenReview ↗

A Gentle Follow-up (Submission #12155)

Dear Reviewer ss4q,

Before the discussion period concludes, we would like to take this opportunity to thank you for your thorough review and valuable feedback. We would be grateful if you could share whether our rebuttal responses and supplementary results have addressed your concerns. We fully understand your time constraints, and we would be happy to address any remaining questions if needed.

Thank you once again for your time and contribution to the review process.

Best regards,

Authors

Reviewer_D6ySOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies incomplete multi-view clustering and proposes Anchor-induced Multi-Granularity Consensus (AMGC). To establish consistent cross-view references despite missing observations, it selects shared-index anchors and aggregates masked sample-anchor assignments without explicit view completion. Anchor graphs are spectrally partitioned at multiple resolutions, and the resulting group relations are transferred to samples. The model further combines sparse local geometry with adaptive low-rank consensus learning. Experiments on five benchmarks under four missing-view rates include comparisons with nine baselines, ablation studies, parameter analyses, efficiency measurements, and supplementary diagnostics. Across the tested datasets and conditions, the results are generally competitive; however, the missingness protocol and interpretation of multi-granularity weighting require clarification.

Critical strengths

1.The motivation is well grounded. The paper identifies two relevant challenges in incomplete multi-view clustering: inconsistent relational references across partly observed views and the varying reliability of structural cues at different scales. Both challenges clearly inform the design of its proposed framework. 2.The method has meaningful technical merit. Shared sample-index anchors provide comparable relational coordinates across views, while spectral anchor partitions transfer different grouping resolutions into the sample space. Integrating mutual-kNN neighborhood geometry and an adaptive low-rank consensus representation offers a coherent mechanism for balancing local and global structural evidence under missing observations. 3.The experimental evaluation is reasonably comprehensive. Table 1 includes nine baselines across five benchmark datasets and four missing-view rates. Table 2 assesses the contributions of individual modules, while the appendix reports optimization behavior, computational efficiency, adaptive-weight diagnostics, and an anchor-selection comparison under class-dependent missingness.

Critical weaknesses

1.The missingness protocol may overstate robustness to limited cross-view overlap. Appendix B.2 retains fully observed samples, and Table 7 reports zero missing anchor-view observations for AMGC. Additional experiments where anchors are incomplete or overlap is severely restricted would strengthen the central empirical claim. 2.Equation (19) favors sources with smaller low-rank fitting errors. Appendix A.3 notes that relations with c_g ≤ K can be individually fitted exactly, irrespective of semantic quality. Please clarify whether the adaptive weights measure structural reliability rather than rank compatibility. 3.Equation (7) independently derives spectral partitions at each granularity; no nesting constraint is imposed. What distinguishes this from multiresolution partition ensembling?

Questions for authors

Refer to the weaknesses.

1 comment
AuthorsOpenReview ↗

A Gentle Follow-up (Submission #12155)

Dear Reviewer D6yS,

As the discussion period draws to a close, we would like to sincerely thank you for your thoughtful comments and constructive feedback. We would greatly appreciate it if you could kindly let us know whether our rebuttal and additional experiments have adequately addressed your concerns. If any questions remain, we would be more than happy to provide further clarification.

Thank you again for your valuable time and efforts.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.