acceptodds
Withdrawn from ICLR 2027

TCRIMVC: Target-Conditioned Retrieval for Recovery and Refinement in Incomplete Multi-View Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Incomplete multi-view clustering (IMVC) aims to discover coherent cluster structures from multi-view data with partially missing observations, yet existing approaches often rely on dedicated cross-view predictors or generative models to infer unavailable representations, thereby introducing additional modeling assumptions and exposing downstream clustering to potential recovery errors. We revisit this problem from a retrieval perspective: relevant semantics may already exist among observed representations and can be reused as evidence for recovery. To this end, we propose a target-conditioned retrieval framework for incomplete multi-view clustering. Given the observed views of an instance and the identity of a requested view, TCRIMVC retrieves relevant latent evidence for representation construction. Missing-view representations are constructed by aggregating retrieved evidence, whereas observed representations are refined through a residual update that retains their original features. Both pathways retrieve from encoded observations, without reusing completed outputs as evidence. The resulting representations are jointly optimized through view reconstruction, global-to-view alignment, and confidence-aware balanced cross-view consistency, reducing the influence of uncertain recoveries on consensus learning. Extensive experiments on diverse benchmarks demonstrate the effectiveness and robustness of TCRIMVC under varying incomplete conditions. Further analyses validate the contributions of its retrieval, refinement, and consensus mechanisms, while cross-view image completion results illustrate the broader applicability of the proposed retrieval principle.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Representation Learning, Missing-View Representation Recovery

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

The authors presented a retrieval-based framework (TCRIMVC) that reformulates incomplete multi-view representation learning as target-conditioned direct-context retrieval. Experiments across multiple datasets and missing rates demonstrate the effectiveness and robustness.

Critical strengths
  1. Reframing missing-view recovery from conventional prediction or generation to target-conditioned evidence retrieval is somewhat innovative.

  2. Modulating cross-view clustering consistency via retrieval confidence can, to some extent, mitigate the impact of unreliable completions on clustering results.

  3. The experimental analysis is relatively comprehensive and supports the model design from multiple perspectives.

Critical weaknesses
  1. The paper constructs the query using the target-view embedding and further incorporates a view-relation matrix before performing Top-K retrieval. Does the performance gain actually come from target conditioning, or merely from ordinary similarity-based retrieval?

  2. How is the accuracy of missing-view completion measured? It appears that the model may simply bring the representations of same-class samples closer through retrieval, while the existing metrics (ACC、NMI、ARI) are only valid for clustering performance.

Questions for authors

See Weaknesses

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

This paper proposes TCRIMVC, a target-conditioned retrieval framework for incomplete multi-view clustering, inspired by the “retrieve before infer” idea of RAG. Incomplete multi-view inputs are mapped onto a shared unit hypersphere, and missing-view recovery and observed-view enhancement are handled in a unified way using separate recovery and enhancement evidence banks, without introducing a separate generator. The motivation of the method is reasonable, the figures are well drawn, and the experiments are fairly sufficient, with comparisons against 10 methods.

Critical strengths
  1. The motivation is clear and the method is well defined and easy to follow.
  2. The experimental results are good, and the method can effectively recover the missing views.
Critical weaknesses
  1. The most relevant predecessors are missing from the comparison. DSIMVC (ICML 2022), ProImp (IJCAI 2023), and MICA (Neural Networks 2025) are the closest neighbor-based and attention-based imputation methods, yet only SCVT appears in Table 1. The retrieval module is essentially a sparse cross-attention over observed latents, so the comparison with ProImp in particular is necessary to establish the claimed novelty. A trivial baseline that fills missing slots with the mean of the instance's own observed views and uses the same losses should also be included.
  2. The recovery path may degenerate into a weighted average of the instance's own observed views. The query in Eq. 3 is built from the mean of the instance's observed views, the evidence bank in Eq. 2 contains those same views, and Eq. 4 adds a same-instance bias that, as the authors note, can only favor them. Since the values are not transformed toward the target view and the recovered representation lies in the convex hull of the selected latents (Prop. 2), the "view v" representation is just an average of latents mostly drawn from other views, which is then mean-fused in Eq. 10 and aligned to the fusion in Eq. 12. The whole pipeline could be close to imputation-free mean fusion with mild neighbor smoothing. The paper provides no statistics on which entries are actually retrieved and no ablation on the same-instance bias or on restricting candidates to other instances or to the target view.
  3. The missing-rate protocol is underspecified. The paper never gives a mathematical definition of η, the mask sampling procedure, or the resulting missingness statistics. This matters because Table 1 and Figure 2(a) show the counterintuitive pattern that performance at higher missing rates sometimes exceeds the complete-data setting.
Questions for authors
  1. The datasets MNIST-10k, Reuters-1200, and NGs come from the classical graph-based multi-view clustering literature, and none of the deep baselines report results on them in their original papers (for example, DIMVC-HIA uses BDGP, MNIST-USPS, Fashion, and Handwritten; DGIMVCM uses HW, Scene-15, LandUse-21, and 100Leaves). Why did the authors not adopt the standard benchmarks of the deep baselines? Under the current suite all baseline numbers are re-implementations, and several of them are close to chance level on NGs and Reuters-1200, which raises concerns about tuning and preprocessing.
  2. In Figure 2(a) and Table 1, accuracy at η = 0.5 is higher than at η = 0 on several datasets. If retrieval-based recovery acts as a regularizer, this should be stated and analyzed as a finding. If it stems from a protocol difference between the complete and incomplete settings, the η = 0 comparison is not fair. Please give a precise definition of η and clarify whether the training procedure is identical across settings.
Reviewer_a8wxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TCRIMVC, which recovers missing views through target-conditioned retrieval, enhances observed views through residual retrieval, and improves clustering performance through reliability-aware cross-view consistency learning. The experimental results demonstrate the effectiveness of the proposed method.

Critical strengths

This paper proposes a target-conditioned retrieval framework for incomplete multi-view clustering, which unifies missing-view recovery and observed-view enhancement within a single retrieval mechanism. It further employs reliability-weighted cross-view consistency learning to mitigate the impact of erroneous imputation. The experiments are relatively comprehensive, and the method achieves good clustering performance across multiple datasets.

Critical weaknesses
  1. The main components introduced in this paper include target-conditioned queries, Top-K weighted aggregation, and residual enhancement. However, their fundamental differences from existing neighbor-based imputation and attention-based aggregation methods remain unclear.

  2. The contributions of the target-view embedding, cross-view relation matrix, and same-sample bias have not been evaluated separately, making it difficult to justify the necessity of the target-conditioned retrieval design. The existing experiments only remove the entire CVC loss, lacking a direct comparison with standard equally weighted CVC.

  3. On MNIST-10k, Reuters-1200, and Scene-15, the clustering performance at a 50% missing rate is even better than that under the complete-data setting. The authors should explain this phenomenon.

  4. The largest dataset used in the experiments contains only 10,000 samples, whereas traditional MVC methods have already been evaluated on datasets with hundreds of thousands of samples, such as YoutubeFace. I suggest evaluating the proposed method on larger datasets.

Questions for authors

I would appreciate the authors addressing the concerns raised in Critical Weaknesses. I will consider revising my score based on their responses.

Comments on the paper

1 comment

Dear Reviewers and Authors,

As a quick reminder, the author-reviewer discussion period will close on November 18.

@Reviewers: Please make sure to read the authors' rebuttals, acknowledge their responses, and engage in further discussion if any points remain unclear. While final decisions are not made at this stage, we ask that you carefully re-evaluate your ratings in light of the new clarifications and experimental results:

Increase your score if the authors have satisfactorily resolved your concerns and strengthened the paper.

Decrease your score if new, unaddressed issues have emerged during the discussion.

Maintain your score if critical concerns remain unresolved.

Please note that you are under no obligation to change your rating, provided your final score accurately reflects the current state of the manuscript. To ensure a fair evaluation, please ground your assessment in concrete evidence rather than personal preference, and consider whether the work offers meaningful scientific value to the ICLR community.

Thank you very much for your time, thoughtful engagement, and contributions to the review process.

Best regards,

Area Chair

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

Related papers

Discussion (0)

Sign in to comment.