acceptodds
Withdrawn from ICLR 2027

Complete-Sample-Driven Shared Semantic Prototype Learning for Highly 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Withdrawn from ICLR 2027, so left out of the paper lists and the map.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 latent cluster structures from multi-source data with missing views. However, under high missing rates, the limited cross-view correspondences often lead to unstable representations, semantic misalignment across views, and biased cluster prototypes. To address these issues, we propose a complete-sample-driven framework with shared semantic prototypes, denoted as CSSP, for highly incomplete multi-view clustering. Specifically, a small set of complete samples is first used to pretrain view-specific autoencoders for stable representation initialization. During prototype training, artificial missing-view tasks are constructed by randomly masking complete samples, and batch-wise missing-view consistency regularization is imposed on latent representations, reconstructions, and prototype assignments, enhancing the model’s adaptability to diverse missing patterns. Furthermore, shared semantic prototypes are constructed from complete samples to provide a reliable global semantic reference across views. Shared assignment consistency and prototype-level soft bridging are then introduced to align view-specific prototypes with the shared semantic space. Finally, the learned view prototypes are used to recover missing latent representations for clustering. The proposed method effectively exploits reliable cross-view semantics from a limited number of complete samples without relying on complex generative view recovery. Extensive experiments on multiple multi-view datasets under various high-missing-rate settings demonstrate competitive clustering performance and strong robustness, particularly in highly incomplete scenarios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering, high missing rate, meta-learning regularization, shared semantic prototypes, prototype-level soft alignment

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 4 reviews, above 27% of the others.
Reviewer_W9pdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an incomplete multi-view clustering method, named CSSP. Specifically, CSSP introduces objectives to learn cross-view consistency. After that, CSSP constructs semantic prototypes from complete samples to guide cross-view prototype alignment. Finally, the aligned prototypes are used to recover missing-view representations for further multi-view clustering. Experiments on multi-view benchmarks demonstrate competitive performance of CSSP compared with existing SOTA methods.

Critical strengths
  1. The paper addresses an important problem in incomplete multi-view clustering, namely learning cross-view consistency when most views are missing.

  2. Experiments on six datasets demonstrate the effectiveness of the proposed CSSP.

  3. The paper is well structured and supported by extensive ablation studies.

Critical weaknesses
  1. My main concern is the novelty of the proposed CSSP. To be honest, I do not see the differences between this work and previous studies. The proposed CSSP appears to be an incremental method that simply combines existing techniques, e.g., view-specific reconstruction, cross-view consistency, prototype learning, and missing-view recovery. I think the authors should clarify the novelty of the designed CSSP.

  2. I am also concerned about the motivation of the claimed problem, i.e., highly incomplete multi-view clustering. Related works, such as Semantic Invariant Multi-view Clustering with Fully Incomplete Information (IEEE TPAMI), have investigated multi-view clustering under severe incomplete settings, where no complete samples are available. Therefore, I agree that highly incomplete multi-view clustering is not a novel problem. As the methodological novelty also remains unclear, I do not believe the paper meets the standards for acceptance at ICLR.

  3. Besides, the experiments focus on high missing settings, while lacking evaluations at various missing rates, e.g., 10%, 30%, 50%, and even 100% missing views. Without the corresponding results, it is difficult to verify the effectiveness of the proposed CSSP across a broader range of missing-view settings.

  4. The motivation for each component is insufficient. The authors should clarify the reasons for proposing each component and explain why it is necessary, rather than simply describing its formulation.

Questions for authors

See the weaknesses.

Reviewer_bCS6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses incomplete multi-view clustering under high missing rates and proposes the CSSP framework. It constructs shared semantic prototypes from complete samples and enhances semantic consistency across views through missing-view consistency constraints and a prototype-level soft alignment mechanism. Experiments on six benchmark datasets under different missing rates demonstrate that the proposed method achieves strong clustering performance and robustness.

Critical strengths

The paper addresses scenarios where cross-view correspondences are limited under high missing rates, making it practically relevant. Meanwhile, it introduces a prototype-level soft alignment mechanism that avoids the constraints of strict one-to-one matching, and the ablation study confirms its contribution to clustering performance.

Critical weaknesses
  1. The object called a shared semantic prototype is, as described, one representation for each complete sample rather than a learned cluster representative. This is more than a naming issue: the bank’s size and coverage are determined by the surviving complete subset, so the method may be sensitive to redundant or unrepresentative references. The paper should state this distinction explicitly and show whether a smaller, clustered bank gives similar behavior.

  2. The paper reports only average results and does not account for the randomness introduced by complete-sample selection and model initialization. Without reporting the variability across multiple missingness masks and training runs, it is difficult to determine whether the modest performance gains over the strongest baseline are reliable.

  3. The computational case for CSSP is not yet demonstrated. The paper contrasts its approach with more involved recovery mechanisms, but does not report training time, inference time, or peak memory for comparable implementations.

Questions for authors
  1. When the missing-view patterns are generated, how many samples remain complete at each missing rate, and is the same complete subset used across methods and random seeds? Because this subset drives both autoencoder pretraining and construction of the semantic bank, its composition could materially affect the comparison.

  2. Is the shared semantic bank fixed after the autoencoder pretraining stage, or is it updated jointly during prototype learning? Please also clarify which losses send gradients to the bank and the exact rule used to recover a missing latent representation from it.

  3. What happens if the number of reference samples is capped while the dataset size changes? A small subsampling experiment, ideally including a clustered version of the bank, would help separate the effect of reference-set size from the proposed consistency mechanism.

  4. For the closest baselines, including DCG and PMIMC, please report variability across missingness masks and training seeds together with wall-clock time and peak memory. An appendix table would be sufficient.

Reviewer_7QHxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates incomplete multi-view clustering (IMVC) under high missing view rates. Their proposed method, CSSP, consists of a three-stage pipeline that makes use of a small set of complete samples. The complete samples are used to pretrain view-specific autoencoders, then randomly masked views to artificially construct missing-view tasks to learn latent representations, reconstructions, and prototype assignments with batch-wise missing-view consistency regularization. Prototypes are learned from the complete and artificially constructed incomplete samples, and missing views are recovered via prototype matching. Importantly, CSSP enforces semantic alignment between complete and missing views and improve representation stability by optimizing for several consistency losses. CSSP achieves competitive results compared to other methods across six benchmark datasets with high missing view rates (0.7-0.95).

Critical strengths

This paper addresses an important problem of learning under multi-view settings with high missing view rates. The problem is well-motivated and defined, and the figures are clear.

Critical weaknesses

Overall, the paper would benefit from some reorganization for better readability. For example, qi(v)q_i^(v) in line 161 is defined in Eq. (11) on line 224, which can be difficult to follow at first glance. More intuition on the design of each loss and how they interact would also be helpful.

  • There is no explicit mention of the pretraining stage in the paper outside of “view-specific autoencoders”, or what Lpre\mathcal{L}_{\text{pre}} is (see Figure 1). The encoder and decoder formulations should be specified. It is also unclear how the complete samples are chosen, and how many are needed to obtain “stable” representations.
  • Typo in Figure 1: Lmvc\mathcal{L}_{\text{mvc}} instead of Lmeta\mathcal{L}_{\text{meta}}
  • Eq. (9): Is simsim cosine similarity?
  • Experiments: Missing models and training details, such as learning rate, batch size, optimizer, and how the temperature (τ\tau) in Eq. (10), (11), and tt in Eq. (10) are selected. The performance of the model remains unconvincing, as SCVT underperforms similar methods, such as PMIMC and GMAE, on a variety of datasets and missing rates.
  • CSSP does not win in many cases. Please provide an explanation of when and why.
  • Section 3.4: Why is λmvc\lambda_{\text{mvc}} not explored here? How does each term in Eq. (5) affect performance?
  • Figure A.1 is misleading, as multimodal data is not explored in this work. However, it would be interesting to see how SCVT performs on representations derived from data with multiple modalities.
  • Many key details and figures in the Appendix, such as the core CSSP algorithm, could be included in the main paper for clarity.
Questions for authors
  • Please addresses the weaknesses above.

Additional questions:

  • If missing data is determined at random, can you run the experiments across different random seeds? Furthermore, error bars are missing from the experimental results.
  • Do all the competing methods utilize a set of complete samples as well?
Reviewer_fCLzOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a novel incomplete multi-view clustering method. This method utilizes a small set of complete samples that retrain cross-view semantic correspondences. Then these samples provide shared semantic information and are randomly masked to construct artificial missing-view tasks.

Critical strengths

1.The idea of this paper is clear.

Critical weaknesses
  1. This paper adopts random masking on a small number of complete samples to generate data of missing views, which raises a critical issue. Specifically, the manuscript does not draw a conclusion regarding how many samples are required to produce high-quality missing-view data, i.e., the sample size sufficient to faithfully recover the underlying data distribution. Corresponding justifications should be provided.

  2. In Tables 1 and 2, the proposed method does not attain state-of-the-art performance for the majority of settings, further weakening the evidence for its effectiveness.

  3. The ablation studies in Table 3 fail to reveal performance improvement trends for each proposed module, which further demonstrates a critical flaw of the method: instability. This indicates that the claims in the manuscript are overstated.

  4. Given the era of large language models (LLMs), authors are suggested to integrate LLMs into their framework to improve algorithm performance and incorporate LLMs-driven multi-view clustering methods [1] as additional baselines.

[1] LLM-DAMVC: A Large Language Model Assisted Dynamic Agent for Multi-View Clustering

Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.