acceptodds
Under review as a conference paper at ICLR 2027

INCOMPLETE MULTI-VIEW CLUSTERING VIA REP- RESENTATION RECOVERY AND DIRECT SOFT- ASSIGNMENT ESTIMATION

Abstract

Incomplete multi-view clustering (IMVC) aims to discover latent cluster struc- ture from multi-view data with partially observed views. Existing imputation- based methods commonly recover missing representations from observed views and then obtain cluster assignments from the completed representations. For a sample missing a target view, however, representation recovery and soft cluster- assignment estimation generally do not share the same reference samples and re- lation weights, so the latent representation and cluster-membership information already available in the target view are not used under a common sample relation. We propose ReSA, an IMVC method based on representation recovery and direct soft-assignment estimation. ReSA first learns cross-view relation representations for reference retrieval and relation-weight computation. Cross-view contrastive learning treats the relation representations of the same jointly observed instance as a positive pair and those of different instances as negative pairs; neighborhood distribution consistency further preserves local cross-view relations. For a sam- ple with a missing target view, ReSA selects the most strongly related references from samples observed in both the source and target views and computes their relation weights. The same references and weights are then reused to aggregate target-view latent representations for missing-representation recovery and target- view soft assignments for direct assignment estimation. Completed multi-view representations provide global clustering information, whereas observed or esti- mated view-specific assignments preserve view-specific cluster membership; both are integrated by global–view consensus clustering. Experiments on four IMVC benchmarks and multiple missing rates show that ReSA achieves leading cluster- ing performance. Comparisons with the trained model held fixed show that direct soft-assignment estimation is closer to the reference assignment computed from the complete target view under identical references and weights, and missing-view reconstruction confirms that the recovered representations retain decodable target- view content.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
incomplete multi-view clustering, representation recovery, cluster assignment, consensus clustering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LVurOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This manuscript points out that in IMVC, existing imputation-based methods mostly recover missing views, and estimate cluster assignments in separate steps. Therefore, the authors propose to use the same reference samples and relation weights for both view recovery and direct soft assignment throughout the whole procedure. The proposed method, ReSA, first derives the nearest neighbors as reference samples, and uses them for two purposes: (1) recovering missing views and aggregating all views to obtain a global assignment; (2) getting directly soft-assignments. All results are ultimately integrated through consensus clustering. Extensive experiments demonstrate the effectiveness of this method.

Critical strengths

1.The authors have proposed a well-designed method which achieves strong performance on a wide range of datasets.

2.The manuscript is clearly written and technically sound. The authors identify an under-explored design choice and propose a novel approach based on shared references and relation weights.

Critical weaknesses

Despite my appreciation of the writing and reasoning, the significance of this manuscript is not sufficiently established. The authors do not sufficiently justify the benefits of sharing references and relation weights over using separate relations for the two estimation tasks. It seems also natural to obtain missing views from one set of references and cluster with another. Moreover, some methods adopt different approaches for view recovery and clustering, making it unclear whether sharing the same references would be appropriate or beneficial in these settings.

Questions for authors

1.Why is not sharing references a problem? I would like to see further clarifications on this.

2.A key design of ReSA is to integrate assignments from both global assignment and view-specific assignments. I wonder if this framework would work even without shared references and weights. Have the authors considered using different reference sets or relation weights for representation recovery and direct soft-assignment estimation, while retaining the same global–view consensus clustering procedure?

3.In the global assignment branch, representation recovery and global assignment estimation do not directly share the same references and weights. Therefore, this branch alone still follows a paradigm similar to the one criticized in the motivation. Why does ReSA introduce sharing between two parallel branches rather than between the two sequential stages of representation recovery and assignment estimation?

4.As a minor reminder, the AI use disclosure appears incomplete. The ICLR 2027 AI Policy for Authors provides an example that also explicitly states which tasks requiring disclosure did not involve AI use. I encourage the authors to revise this section in accordance with the ICLR 2027 AI Policy for Authors.

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

This paper studies missing-view representation recovery and cluster-assignment estimation in incomplete multi-view clustering (IMVC). It proposes ReSA, which learns cross-view relation representations through cross-view contrastive learning (CCL) and neighborhood distribution consistency (NDC). For a sample with a missing target view, ReSA retrieves reference samples observed in both the source and target views and reuses the same references and weights to aggregate the target-view latent representation and soft cluster assignment separately. A global–view consensus stage combines clustering information from completed representations with view-specific assignments. Experiments on four benchmarks at three missing rates, together with ablation, assignment-estimation, and missing-view reconstruction analyses, evaluate the method.

Critical strengths

(1)The paper identifies a meaningful gap between recovering a missing representation and estimating its cluster assignment. Reusing the same reference samples and relation weights for both tasks is intuitive and provides a consistent sample-level design. The overall pipeline is easy to follow.

(2)As reported in Table 2, ReSA achieves the best mean results across the evaluated datasets, missing rates, and clustering metrics. Its advantage remains visible at higher missing rates, and the paper reports means and standard deviations over three training seeds. These results provide substantial empirical support for the proposed approach.

(3)The ablation study examines relation representations, direct soft-assignment estimation, and global–view consensus. The additional Hellinger-distance comparison and paired Wilcoxon test support direct assignment estimation under the controlled comparison, while reconstruction experiments on Carl and Multi-Modal CelebA-HQ help show that recovered representations retain decodable target-view content.

(4)The paper describes the model architecture, optimization settings, missing-rate protocol, and evaluation metrics. Using shared observation masks across compared methods also helps make the main benchmark comparison more controlled.

Critical weaknesses

(1)Equations (3)–(6) define CCL and NDC over jointly observed samples, while the implementation section gives a mini-batch size of 256. Specifying whether the relation distributions are computed globally or within each mini-batch would make the method easier to reproduce. This appears to be a clarification issue rather than a concern with the overall idea.

(2)The method uses dataset-specific values of k and fixes the consensus settings to the three largest assignment components and γ = 0.5. A compact sensitivity analysis for k and the consensus settings, or a brief rationale for these choices, would help establish that the reported gains are not overly dependent on a particular configuration.

(3)The related-work section discusses recent methods such as ARSL-IMVC, JVILG, and MICA, but the main table does not include them. A clearer explanation of the technical differences, and additional comparisons where feasible, would help readers better understand ReSA’s contribution. Given the strong results against the included baselines, I view this as an opportunity to improve positioning rather than a fundamental weakness.

Questions for authors

(1)Equations (3)–(6) define CCL and NDC over jointly observed samples, while the implementation section gives a mini-batch size of 256. Specifying whether the relation distributions are computed globally or within each mini-batch would make the method easier to reproduce. This appears to be a clarification issue rather than a concern with the overall idea.

(2)A compact sensitivity analysis for k and the consensus settings, or a brief rationale for these choices, would help establish that the reported gains are not overly dependent on a particular configuration.

Reviewer_jt3HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReSA for incomplete multi-view clustering. It learns view-specific latent representations and cross-view relation representations using reconstruction, contrastive learning, and neighborhood distribution consistency. For a sample with a missing target view, the method retrieves jointly observed reference samples and uses the same neighbors and weights to recover the target representation and directly estimate its soft cluster assignment. A global–view consensus procedure integrates assignments from the completed representations with observed or estimated view-specific assignments. Experiments on four clustering benchmarks, component ablations, controlled assignment comparisons, and auxiliary reconstruction evaluations investigate the effectiveness of this design.

Critical strengths

1-Using the same reference samples and relation weights for both missing-representation recovery and soft-assignment estimation gives the two operations a consistent basis. Section 4.3 provides a particularly useful comparison: with the trained model, assignment map, references, weights, and downstream procedure held fixed, direct estimation reduces the mean squared Hellinger discrepancy to the complete-view reference assignment from 0.05694 to 0.05039. The reported paired test provides additional evidence that the proposed estimator improves assignment approximation under this controlled setup.

2-Table 2 reports the highest mean performance across all 36 dataset–missing-rate–metric combinations, including substantial improvements over competitive baselines in several settings. The evaluation covers datasets with different scales, numbers of views, and numbers of classes. Reporting variation across three training seeds and sharing inputs and observation masks across methods are useful steps toward controlled comparisons. The auxiliary reconstruction experiments also provide complementary evidence that the recovered representations retain target-view content.

Critical weaknesses

1-The main distinction is the reuse of the same reference samples and relation weights for both representation recovery and soft-assignment estimation. This is a sensible refinement, but the conceptual advance over existing approaches that combine imputation and assignment learning seems relatively narrow. The paper explains how the components fit together, while providing less insight into when sharing the relation between these two tasks is particularly advantageous.

2-Table 3 shows ACC improvements of 0.17, 0.63, and 0.26 percentage points on ALOI-100, HandWritten, and Hdigit, respectively, when direct estimation replaces the feature-first alternative. The 3.69-point gain on YouTubeFace10 is more substantial, so the benefit is clearly dataset-dependent. However, the table averages over missing rates and provides no uncertainty estimates for the ablation differences. The controlled Hellinger-distance comparison is useful evidence that direct estimation better approximates the complete-view reference assignment, but the reported significance test does not establish the stability of the final ACC gains. Results broken down by missing rate, with uncertainty estimates, would help clarify the practical importance of this central contribution.

3-DVIMC obtains approximately 1–2% ACC and near-zero ARI on ALOI-100. On YouTubeFace10 at MR = 0.5, its NMI and ARI are both zero. These outcomes suggest degenerate clustering in some settings and deserve further investigation. Although the authors use public implementations and standardize the inputs and evaluation protocol, this does not fully explain the failures. Dataset-specific configurations, preprocessing details, convergence curves, and cluster-occupancy statistics would help determine whether the results reflect genuine limitations of the method or problems with its configuration and optimization.

4-Section 4.1 states that the three training seeds share a single observation mask for each dataset–missing-rate setting. Consequently, the reported standard deviations capture training variability under one particular missingness realization. This matters because the proposed method depends on jointly observed reference samples, whose availability and composition can change with the mask. Sharing a mask across methods is appropriate, but the comparison should also be repeated over several independently sampled masks at each missing rate. The authors should additionally specify how views are removed within incomplete samples, since the same sample-level missing rate can produce substantially different cross-view overlap.

Questions for authors

1-Could the authors report differences in final clustering ACC with uncertainty estimates and explain why the improvement is substantially larger on YouTubeFace10? This would help connect the controlled assignment-discrepancy result to the practical contribution of the central design.

2-Please provide the relevant dataset-specific configurations and diagnostics, including preprocessing, convergence behavior, and cluster occupancy, to clarify whether these failures are intrinsic to the evaluated setting.

3-Please clarify how many views are removed from each incomplete sample and how those views are selected. Repeating the comparisons over several masks, with each mask shared across methods, would directly address the robustness concern.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.