acceptodds
Under review as a conference paper at ICLR 2027

MGRL-IMVC: Multi-Granularity Relational Learning for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to uncover latent cluster structures from partially observed views. However, missing views weaken instance-level cross-view supervision, noisy observations reduce the reliability of available pairs, and cluster-level semantic discrepancies across views hinder consistent representation learning. This paper proposes MGRL-IMVC, a multi-granularity relational learning framework that avoids imputing missing features and jointly models sample relations, neighborhood structures, and prototype correspondences. Specifically, unified-space relational learning aggregates all observable representations and mines latent positive and negative relations, enabling samples without direct cross-view pairs to participate in contrastive learning. Cross-view structural learning constructs affinity-based neighborhood weights to provide structural evidence beyond direct similarity for same-instance pairs and to modulate contrastive constraints. Cross-view soft prototype alignment estimates soft correspondences between view-specific prototypes via the co-occurrence of soft assignments on commonly observed samples, and enforces bidirectional prototype matching for cluster-level semantic alignment. These objectives are jointly optimized with view reconstruction, allowing limited observations to complement each other across multiple relational granularities. Experiments on five benchmark datasets under various missing rates demonstrate that MGRL-IMVC achieves superior clustering accuracy and adaptability compared with seven representative deep IMVC methods, and ablation studies verify the complementary contributions of the three granularities.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Multi-Granularity Learning, Contrastive Learning, Cross-View Structural Learning, Soft Prototype Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Av7kOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Multi-Granularity Relational Learning for Incomplete Multi-View Clustering (MGRL-IMVC). The framework learns from incomplete observations without imputing missing features by integrating three relational granularities. Unified-space relational learning mines potential positive and negative pairs, cross-view structural learning weights contrastive constraints using neighborhood affinity concentration, and soft prototype alignment estimates cross-view cluster correspondences through co-occurrence of soft assignments. Experiments on five benchmarks under different missing rates demonstrate competitive clustering performance.

Critical strengths
  • The paper explores multi-granularity relational learning for incomplete multi-view clustering, integrating sample-level relations, neighborhood structures, and cluster-level semantics without explicit missing-view imputation.

  • The soft prototype alignment mechanism uses commonly observed samples to estimate cross-view prototype correspondences, providing a reasonable approach to addressing inconsistent cluster assignments across views.

  • The experiments cover five datasets under different missing rates. The ablation studies and visualizations provide empirical evidence for the benefits of combining multiple relational objectives.

Critical weaknesses
  • The structural reliability score (W_{ii}) in Eq. (6) measures the concentration of a normalized affinity distribution rather than the correctness of cross-view relations. Incorrect but concentrated affinities can receive equally high weights, making its interpretation as relation reliability insufficiently justified.

  • Eq. (7) scales only the positive logit by (W_{ii}), whose effect depends on the sign of the cosine similarity. For negative similarities, decreasing (W_{ii}) can increase the positive softmax probability. This behavior complicates its interpretation as consistently suppressing unreliable relations.

  • The positive and negative relations in Eq. (4) are mined from the evolving unified representation space, where cross-view comparability is not guaranteed during early training. Incorrect relations may therefore be reinforced through subsequent contrastive optimization. No direct analysis of mined-relation quality or its evolution is provided.

  • The prototype alignment mechanism lacks important optimization details, including the K-means refresh schedule and gradient treatment of prototypes and correspondence weights. Moreover, the unnormalized co-occurrence matrix (R) makes the alignment loss dependent on the number of co-observed samples, while Fig. 3 does not directly establish prototype-matching correctness.

Questions for authors
  • How can WiiW_ii distinguish a concentrated affinity on the correct counterpart from an equally concentrated affinity on the wrong sample?

  • How is the reliability-weighting behavior in Eq. (7) justified when the positive cosine similarity is negative?

  • Can the authors report the precision and recall of mined positive and negative relations throughout training to verify that relation quality improves?

  • How frequently are K-means prototypes updated, which quantities receive gradients, and can the authors provide quantitative prototype-matching accuracy?

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

The paper presents a meaningful attempt to address incomplete multi-view clustering through multi-granularity relational learning. The framework is reasonably motivated, and the experimental results demonstrate its potential effectiveness. However, the methodological contribution relative to recent studies remains somewhat incremental, and several key design choices require stronger justification. More comprehensive ablations, clearer optimization details, and additional evidence for the structural reliability and prototype alignment mechanisms would substantially strengthen the paper.

Critical strengths
  1. The paper addresses an important problem in incomplete multi-view clustering and provides a reasonable motivation for exploiting relational information at different granularities.
  2. The proposed framework avoids explicit missing-view reconstruction and jointly considers sample relations, neighborhood structures, and prototype correspondences. The overall framework is clearly presented and relatively easy to follow.
  3. The experimental results are encouraging on several benchmarks, particularly DHA and HandWritten. The ablation studies also demonstrate that the proposed components can provide complementary improvements.
Critical weaknesses
  1. The unified-space relational learning and cross-view structural learning components appear closely related to existing global-local contrastive learning methods, particularly GLGC (CVPR 2026). Meanwhile, cross-view prototype alignment has been explored in previous IMVC studies, such as CPSPAN. The main contribution seems to lie in integrating these existing ideas into a unified framework. The authors should better distinguish their approach from these related methods and explain what new insights are introduced beyond their combination.
  2. The structural reliability estimation in CSL deserves further justification. In Eq. (6), the structural relation matrix is computed as (W=AA^\top), while Eq. (7) only uses its diagonal entries to weight positive pairs. However, (W_{ii}) essentially measures the concentration of the normalized affinity distribution rather than the consistency of neighborhood structures across two views. A concentrated affinity distribution does not necessarily indicate a reliable cross-view correspondence. Could the authors provide a more convincing explanation or additional evidence supporting this design?
  3. The prototype alignment mechanism requires further clarification. The proposed method obtains view-specific prototypes using K-means and constructs prototype correspondence weights from soft-assignment co-occurrence. However, it is unclear how these prototypes are updated during joint optimization and whether gradients are propagated through the correspondence estimation. Moreover, the evidence matrix in Eq. (10) is not normalized by the number of commonly observed samples, which may cause the scale of the alignment loss to vary under different missing rates. These implementation and optimization details should be clarified.
  4. Table 2 mainly considers progressively adding CSL and PA to URL. Additional combinations, particularly URL+PA and variants without URL, would help distinguish the individual contributions and interactions of the three components. Furthermore, Figure 3 visualizes prototype correspondence matrices after reordering the prototype indices. Although the results are visually intuitive, such reordering does not directly demonstrate that the learned representations achieve better semantic alignment. More quantitative evidence would strengthen this claim.
  5. Although the method achieves competitive performance, the improvements are not consistent across all datasets and missing rates, particularly under severe incompleteness. Reporting standard deviations or statistical significance would help assess the reliability of the improvements. In addition, the framework constructs dense sample-relation and cross-view affinity matrices, which may introduce substantial computational and memory overhead. Runtime comparisons, complexity analysis, and clearer implementation details, including how the final clustering results are obtained from incomplete observations, would make the evaluation more convincing.
Questions for authors
  1. What is the main methodological distinction between the proposed URL/CSL components and the global-local relational learning mechanisms in GLGC?
  2. Why is the diagonal of (AA^\top) sufficient to characterize the structural reliability of cross-view positive pairs? Have the authors considered alternative neighborhood-consistency measures?
  3. How frequently are the K-means prototypes updated during training? Are the correspondence weights treated as fixed targets or jointly differentiated?
  4. How does the number of commonly observed samples affect the prototype alignment loss, especially under high missing rates?
  5. Could the authors provide additional details on the final clustering procedure and the computational overhead introduced by the relational learning components?
Reviewer_HSJzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MGRL-IMVC, a multi-granularity relational learning framework for incomplete multi-view clustering that circumvents the need to input missing features, and jointly models sample relations, neighborhood structures, and prototype correspondences. Extensive experiments are conducted to verify the effectiveness of the proposed method.

Critical strengths

S1. Figure 1 provides a clear and intuitive illustration of the overall framework, which facilitates understanding of the method.

S2. Exploiting three granularities to improve the performance of incomplete multi-view clustering is a reasonable and well-motivated design.

Critical weaknesses

W1. Treating top-kk and bottom-kk similarity samples as positive and negative samples, respectively, lacks theoretical justification. The authors should provide more theoretical analysis to support this design choice.

W2. The clustering results in Table 1 show that the proposed method performs worse than the compared methods on the NoisyMNIST and ALOI datasets. This degradation should be discussed in more detail, including its possible causes.

W3. The paper relies heavily on step-by-step descriptions of the pipeline and provides limited theoretical support for the method. The authors are expected to strengthen the theoretical analysis (e.g., the rationale behind the proposed objectives) rather than merely describing the procedural steps.

Questions for authors

Please address the weaknesses in your rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.