acceptodds
Under review as a conference paper at ICLR 2027

SEPro: Semantic Evidence Propagation for Incomplete Multi-View Clustering

Abstract

Incomplete Multi-View Clustering (IMVC) aims to learn coherent representations from data with missing views. However, observed views may also be noisy, biased, or incomplete in semantics, making existing reconstruction- and consistency-based methods vulnerable to unreliable information. To address this issue, we propose SEPro, a semantic evidence propagation framework for IMVC. SEPro represents each view with explicit semantic evidence, uncertainty, and confidence, and propagates this evidence through prototype-mediated high-order relations guided by view availability, semantic affinity, and confidence. The learned evidential representations also support uncertainty-aware inference for missing views. By jointly integrating evidence extraction, propagation, and inference, SEPro learns robust consensus representations for incomplete multi-view clustering. Extensive experiments demonstrate its effectiveness and robustness under varying missing-view conditions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering, Semantic evidence, Uncertainty modeling, Evidence propagation, Missing-view inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_g5ADOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes to unify evidence representation, evidence propagation, and missingview inference into a joint optimization framework. The direct exploitation of uncertainty and confidence is reasonable. The authors conduct experiments on six standard multiview datasets together with comprehensive tSNE visualization analyses.

Critical strengths

Sufficient experimental comparisons are performed against several stateoftheart (SOTA) methods.

Critical weaknesses

1.The proposed framework falls into the category of incremental innovation. 2.Although extensive comparative experiments are carried out, there is still insufficient evidence to demonstrate that the proposed method achieves significant advantages on most datasets. 3.The performance of the method is dependent on quality of the prototype.

Questions for authors

1.It is suggested that the motivation and rationality of the proposed method be elaborated more thoroughly. 2. On the issue that the method’s performance relies on prototype quality, please provide preliminary ideas for future research.

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

This paper studies IMVC. The authors propose SEPro, which models each observed view as semantic evidence with a confidence score, propagates the evidence through prototypes to obtain a consensus representation, and uses the same evidence to infer missing views. Experiments are conducted on six datasets under different missing rates.

Critical strengths
  1. The motivation is reasonable. Observed views are not equally reliable, and the paper tries to model this.
  2. The method is simple and easy to follow. Experiments cover six datasets, and several missing rates show the better performance.
Critical weaknesses
  1. The overall architecture closely follows HSACC (NeurIPS 2025): the same autoencoder backbone, the same feature-level mutual information objective, the same training configuration, and the same evaluation layout. As far as I can tell, the contribution relative to HSACC reduces to (a) replacing its MMD-based weighted fusion with a sample-to-prototype-to-sample weighted average, and (b) replacing its MSE inference head with a Gaussian NLL head plus a scaling factor. HSACC is listed only as a baseline and is never acknowledged or discussed as the basis of the method. The authors should state this relationship explicitly and isolate the gain of each replacement.
  2. The term "evidential" is misleading. In the trusted multi-view learning literature (Han et al. ICLR 2021 and follow-ups) it refers to Dirichlet or subjective logic frameworks, which this paper does not use; what is actually learned is a heteroscedastic Gaussian head. More importantly, the abstract and all three contribution claims rest on confidence-guided propagation, yet the confidence score u receives no direct supervision, has no stated relation to the variance, and is never validated against any measurable notion of reliability (for example, with noisy or corrupted views).
  3. The baseline numbers are not comparable. The DCG results on Handwritten are copied verbatim from the DCG paper, but DCG uses a different missing protocol (randomly selecting samples and deleting one view each), its official code uses only two of the six Handwritten views, and its repeated runs effectively share a single seed. Placing these numbers in the same table as the six-view SEPro results, under a missing protocol that this paper never defines, makes the comparison unreliable.
  4. The SOTA claim does not hold on Caltech-5V. Publicly reported results for DVIMC (Xu et al. 2024) and PMIMC (Yuan et al. 2025) on this dataset are clearly higher than those of SEPro under comparable missing rates. Both methods are cited in the paper but excluded from the comparison without explanation.
  5. The ablations (Tables 2 and 4) only toggle loss terms. None of the mechanisms that define the contribution are isolated: prototype-mediated propagation versus simple averaging or pairwise graph propagation, removing the confidence weight u, removing the 1/(1+σ²) and α calibration in completion, and sensitivity to K and τ. Without these, the claimed benefits of the propagation and the uncertainty-aware completion remain unsupported.
Questions for authors

See the weakness

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

This paper proposes SEPro, a semantic evidence propagation framework for incomplete multi-view clustering. SEPro uses learned semantic evidence and confidence scores to infer missing views. The overall learning framework consists of reconstruction loss, cross-view consistency loss, evidential representation loss, and cross-space semantic loss. These components aim to preserve view-specific information, improve cross-view consistency, learn semantic evidence, and align representations across different views. The experimental results are promising.

Critical strengths
  • Experimental Evaluation. The comparison results, which are reported in Table 1, demonstrate the consistent improvement of SEPro. The detailed ablation study is good. Authors have studied all combinations in Table 2. The results show the contribution of different loss components.
  • Problem Motivation and Framework Design. The proposed framework integrates view availability, semantic affinity, and confidence for evidence propagation, providing an intuitive approach to learning consensus representations from incomplete multi-view data.
Critical weaknesses
  1. The paper uses both predictive variance and confidence scores, but their relationship and distinct roles are not clearly explained. It is also unclear how the confidence scores are learned.
  2. Eq. (9) seems to miss a latent dimension factor in the log-variance term of the Gaussian NLL. Also, Eq. (4) uses cross-view feature products as a joint probability distribution without ensuring non-negativity or normalization.
  3. The interactions between different components are somewhat complicated. A clearer theoretical explanation would help justify the design of the overall framework.
  4. The related work section discusses evidential multi-view clustering methods but lacks relevant references to support the discussion.
Questions for authors
  1. Prototypes. How are the semantic prototypes ckc_k in Eq. (12) initialized and updated during training?
  2. Missing-view inference. Why does Eq. (22) use μiv/(1+σi2(v))\mu_i^v/(1+\sigma_i^{2(v)}) for uncertainty suppression? Reducing the magnitude of a representation does not necessarily improve its accuracy. Could the authors provide an ablation study comparing different scaling strategies? Also, could the authors evaluate missing-view prediction errors by masking observed views?
  3. Evidence aggregation. In Eqs. (16)–(18), the same propagation weights are used twice: first to aggregate evidence into prototypes eke_k, and then to construct the consensus representation hih_i. Could the authors explain why the same weights are used in both steps and provide more justification for this two-stage aggregation?
  4. Novelty. Some existing IMVC methods have used related prototype-based mechanisms. For example, ProImp (IJCAI 2023) uses bidirectional sample-prototype aggregation, while FreeCSL (CVPR 2025) learns consensus representations and shared semantic prototypes. Could the authors further clarify the main differences between SEPro and these methods, particularly beyond the confidence-based weighting?

Please correct me if I have misunderstood anything. I will carefully consider the authors' rebuttal before deciding on my final rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.