acceptodds
Under review as a conference paper at ICLR 2027

Unifying Latent Tensorized Graph Learning and Representation-Induced View Recovery for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to address the challenge of discovering consistent information across heterogeneous multi-view data with missing views, a scenario frequently encountered in real-world applications. Although several remarkable and representative methods have been proposed to recover the missing information, they often suffer from either suboptimal graph construction or inaccurate view recovery, constrained by the lack of well-connected representations that can simultaneously support consistent structural learning and accurate view completion. In this paper, we propose a novel framework, Tensorized Graph learning with Representation-induced Smooth recovery, which unifies latent tensorized graph learning and representation-induced view recovery in a joint optimization scheme. Specifically, the tensorized graph learning captures high-order cross-view structural correlations, while the smooth representation learning guides accurate recovery of missing views through low-rank and manifold-preserving constraints. This synergy enables robust representation learning. We derive efficient closed-form solutions for each subproblem, ensuring monotonic convergence. Extensive experiments on eight real-world multi-view datasets demonstrate that TGRS-IMVC consistently outperforms state-of-the-art IMVC methods across a range of missing ratios. Ablation studies and parameter sensitivity analyses further validate the effectiveness and stability of the proposed components.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
clustering, tensorized graph learning

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_4F7cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TGRS, a unified framework for incomplete multi-view clustering (IMVC), which employs smooth self-representation learning to guide missing-view recovery and tensorized graph learning to capture high-order cross-view structural correlations and promote structural consistency. These components are jointly optimized through an alternating optimization strategy, and the resulting consensus graph is used for spectral clustering. Experiments on eight datasets under three missing ratios show that TGRS outperforms seven IMVC baselines in most evaluated settings, while ablation studies further support the contributions of the main components to overall clustering performance.

Critical strengths
  1. The paper presents a well-motivated joint framework that integrates smooth self-representation learning and tensorized graph learning, enabling mutual reinforcement between missing-view recovery and structural learning.
  2. The method achieves competitive clustering performance across eight datasets and three missing ratios, outperforming seven IMVC baselines in most settings. Ablation studies further support the effectiveness of the two main components.
Critical weaknesses
  1. The mathematical derivations contain several inconsistencies. For example, Eq. (22) does not generally follow from Eq. (19), while Eqs. (24)~(27) contain inconsistencies in the reconstruction loss and representation update, raising concerns about the correctness of the optimization procedure.
  2. The abstract claims monotonic convergence and parameter sensitivity analyses, but neither a convergence proof nor parameter sensitivity results are provided in the main paper.
  3. The effectiveness of representation-induced view recovery is not directly validated, as the experiments only report clustering metrics without evaluating the accuracy of recovered missing views.
  4. The novelty of the proposed framework relative to existing IMVC methods is not sufficiently established. In particular, the contribution of the iterative interaction between view recovery and tensorized graph learning is not clearly isolated or validated.
  5. The related work is incomplete, particularly regarding missing-view recovery methods. Additionally, the description of the two-step graph propagation method does not match the corresponding Zhang et al. (2024) reference, which concerns a different method.
Questions for authors
  1. Could the authors clarify the inconsistencies in Eqs. (19), (22), and (24)–(27)? Are these issues limited to the mathematical presentation, or do they also affect the implemented optimization algorithm and experimental results?
  2. Could the authors provide a theoretical justification for the claimed monotonic convergence and the parameter sensitivity results mentioned in the abstract?
  3. Could the authors directly evaluate the accuracy of representation-induced view recovery using reconstruction error metrics and comparisons with existing recovery methods? This would help verify whether the proposed mechanism improves missing-view recovery beyond the reported clustering performance.
  4. Could the authors clarify the specific novelty of the proposed joint framework compared with existing IMVC methods? In particular, can the benefits of joint optimization be demonstrated through comparisons with a two-stage approach that performs view recovery and graph learning separately?
  5. Could the authors expand the discussion of related work, particularly missing-view recovery methods, and clarify how the proposed approach differs from them? Additionally, please verify the Zhang et al. (2024) reference, which appears inconsistent with the description of the two-step graph propagation method.
Reviewer_nZg6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a novel framework, TGRS (Tensorized Graph learning with Representation-induced Smooth recovery), which unifies latent tensorized graph learning and representation-induced view recovery in a joint optimization scheme. Extensive experiments on eight real-world multi-view datasets demonstrate that TGRS-IMVC consistently outperforms state-of-the-art IMVC methods across a range of missing ratios.

Critical strengths

1.Simultaneously coupling view recovery, representation learning, and tensorized graph learning in one scheme is a sensible design intention and aligns with recent trends in IMVC. 2.Extensive experiments on eight real-world multi-view datasets demonstrate that TGRS-IMVC consistently outperforms state-of-the-art IMVC methods across a range of missing ratios.

Critical weaknesses

1.It is recommended that the authors further summarize the contribution precisely to highlight the novelty of this work. 2.The authors propose a framework that jointly integrates smooth representation learning, tensorized graph learning, and missing view recovery within a single optimization scheme. Please explain how they work together. What is its key technical distinction from existing methods that jointly learn graphs and recover missing views? 3.The related work should further summarize the shortcomings of existing methods and better clarify the distinctions between the proposed TGRS method and existing methods.

Questions for authors

1.What are the performance advantages of the proposed method compared to other state-of-the-art methods? 2.Ablations are conducted only on Leaves, BBC, and ORL. Why were the remaining datasets not included? Do these results support conclusions that generalize across all datasets and missing rates? 3.The generation process for missing ratios r∈{0.1,0.3,0.5} needs to be clarified. Are higher missing ratios such as r = 0.7, 0.9 tested? 4.Can the author elaborate on the experimental details and conduct parameter selection experiments? 5.The paper provides quantitative results but lacks qualitative visualisations.

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

The paper proposes TGRS for incomplete multi-view clustering. It jointly learns view-specific self-representations, recovers missing features, and constructs adaptive graphs. A low-rank tensor constraint captures cross-view structural correlations. The fused graph is then used for spectral clustering. Experiments compare TGRS with seven baselines on eight datasets under three missing ratios. Performance is evaluated using ACC, NMI, and Purity, with additional ablation studies on the main components.

Critical strengths
  1. The combination of missing-feature recovery and adaptive graph learning is well motivated. Self-representation learning connects these components, while low-rank tensor regularization captures structural correlations across views.
  2. The reported clustering results are strong across eight datasets and three missing ratios. Comparisons with seven baselines and the ablation results provide empirical support for the proposed combination.
Critical weaknesses
  1. The mathematical derivation is not internally consistent, making it difficult to verify that the updates solve the stated objective.
  1. Eq. (6) imposes the exact reconstruction constraint X = XZ and includes a Frobenius-norm penalty on Z. However, Eq. (18) and Eq. (24) replace the constraint with a squared reconstruction loss, omit the penalty, and introduce a parameter λ. These changes are not explained, so the connection between the original objective and the subsequent subproblems remains unclear.

  2. The derivation of the Z update also needs correction. Eq. (25) uses Tr(ZLsZT), whose gradient involves ZLs when Ls is symmetric. Eq. (26) instead uses LsZ. If the intended regularizer is Tr(ZTLsZ), this should be stated consistently with the definition of the sample representations.

  3. In addition, Eq. (19) defines B=(I-Z)(I-Z)T, so Bmm should be constructed from the corresponding rows of I-Z. Eq. (22) explicitly uses columns instead. The claimed positive definiteness is also not guaranteed by the Gram-matrix construction, which only ensures positive semidefiniteness. Consequently, the inverse required by Eq. (23) may not exist. The authors should reconcile these expressions and explain the conditions under which the updates are valid.

  1. The supplementary material mentioned in the paper was not available to me, which limits my assessment of reproducibility. The reproducibility statement indicates that source code and additional experimental details are provided in the supplementary material. Without access to these resources, I could not check whether the implementation follows the stated objective or clarify the parameter settings and ablation variants. The authors should provide access to the supplementary material so that these points can be verified.
Questions for authors
  1. Which objective is actually implemented in the code? Please explain how the constraint X = XZ, the penalty on Z, and the parameter λ are handled across Eq.(6), Eq.(18), and Eq.(24).

  2. Is the intended graph regularizer Tr(ZLsZT) or Tr(ZTLsZ)? Please reconcile Eqs.(25)- (27) and clarify whether sample representations correspond to rows or columns of Z.

  3. How is Bmm constructed in the implementation, given the different expressions in Eq.(19) and Eq.(22)? What guarantees its invertibility, and how are singular cases handled in Eq.(23)?

  4. Could the authors provide access to the supplementary material referenced in the reproducibility statement, including the source code, parameter settings, and definitions of the ablation variants?

  5. Please correct the typesetting errors and unresolved equation references, such as “Eq. (3)??” and “Eq.equation6,” and remove the stray word “paragraph.” at the end of the abstract. Please also check that all mathematical symbols and cross-references are displayed correctly throughout the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.