acceptodds
Under review as a conference paper at ICLR 2027

Dual Kernel Contrastive Learning for Partially-Aligned Multi-View Clustering

Abstract

Although multi-view clustering has achieved significant progress, most existing methods assume that all views are fully aligned. In real-world scenarios, however, temporal asynchrony, spatial misalignment, and heterogeneous acquisition condi- tions often lead to partially aligned views, posing substantial challenges to effec- tive clustering. Existing partially-aligned multi-view clustering methods mainly suffer from two limitations: they either insufficiently exploit abundant unaligned samples or rely on cumbersome and error-prone explicit sample-level realign- ment, which may lead to suboptimal clustering performance. To address these issues, we propose a simple yet effective method, termed Dual Kernel Contrastive Learning for Partially-Aligned Multi-View Clustering (DKLPVC). Specifically, DKLPVC first exploits cross-view semantic similarity matrices computed from aligned samples to alleviate view distribution shifts and further uses them as prior information to construct a kernel alignment graph. This graph serves as a seman- tic bridge between sparse aligned pairs and abundant unaligned samples, thereby avoiding explicit sample-level realignment. Under the guidance of the learned graph, DKLPVC performs kernel contrastive learning at both the feature and clus- ter levels in a kernel-induced Hilbert space. The feature-level objective enhances local cross-view semantic consistency, while the cluster-level objective corrects local semantic bias from a global clustering perspective by encouraging intra- cluster compactness and inter-cluster separation. Extensive experiments on eight benchmark datasets demonstrate that DKLPVC consistently outperforms existing methods in partially-aligned multi-view clustering tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-View Clustering;Kernel Contrastive Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_pxVxOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a simple yet effective dual kernel contrastive learning method for partial multi-view clustering, which avoids the complex traditional sample-to-sample realignment while alleviating the distribution shift of view features across different views. The authors also provide theoretical analysis to explain the feasibility of the proposed method, and extensive experiments further verify its effectiveness. Extensive experiments on eight benchmark datasets demonstrate that DKLPVC consistently outperforms existing methods in partially-aligned multi-view clustering tasks.

Critical strengths

(1)This work is effective for handling of partially unaligned views. The proposed method has low complexity, strong performance, comprehensive experimental validation, and a solid theoretical foundation. (2)This paper has a clear motivation, a simple yet effective method, and a detailed introduction to the proposed approach.

Critical weaknesses

(1)This paper could further explain the generalization ability of the proposed model, for example, whether the method is also effective for other tasks. (2)This paper only provides an analysis of the time complexity. It would be helpful to also provide an analysis of the space complexity to further demonstrate the efficiency of the model.

Questions for authors

1.Would the method still perform reasonably well on other tasks, such as multi-view partially view-aligned classification?
2.Why did the authors use a Gaussian kernel? 3.The adaptive threshold is the advantages of this design?

Reviewer_3FPxOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limitations of existing partially-aligned multi-view clustering methods, which either insufficiently exploit unaligned samples or rely on cumbersome and error-prone explicit sample-level realignment, by proposing a simple yet effective method termed Dual Kernel Contrastive Learning for Partially-Aligned Multi-View Clustering (DKLPVC): , DKLPVC then performs kernel contrastive learning at both the feature and cluster levels in a kernel-induced Hilbert space, where the feature-level objective enhances local cross-view semantic consistency and the cluster-level objective corrects local semantic bias from a global clustering perspective by encouraging intra-cluster compactness and inter-cluster separation; extensive experiments on eight benchmark datasets demonstrate that DKLPVC consistently outperforms existing methods in partially-aligned multi-view clustering tasks.

Critical strengths

(1)The paper accurately identifies two specific limitations of existing methods: insufficient exploitation of abundant unaligned samples and reliance on cumbersome and error-prone explicit sample-level realignment. It then designs solutions targeting these two issues, with clear motivation. (2)The method is simple yet ingenious in design, avoiding explicit realignment and bypassing error-prone explicit sample-level realignment; the idea is simple yet effective, reducing the method's sensitivity to noisy pairs.

Critical weaknesses

(1)The authors could further clarify how the parameters were selected in the experiments and how multiple parameters were jointly combined and optimized. (2)The method needs to construct an alignment graph in a kernel-induced Hilbert space and simultaneously perform contrastive learning at two levels, so its time and space complexity may be higher than that of single-level contrastive methods.

Questions for authors

(1) Could the authors explain what advantages the Hilbert space offers compared with the Euclidean space?
(2) On large-scale datasets, has any consideration been given to reducing time and space consumption, and how can the method be generalized to other misalignment tasks, such as partially misaligned classification and semi-supervised classification?

Reviewer_eMFJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work tackles the problem of partially view-aligned multi-view clustering, in which cross-view correspondences are available for only a portion of the samples and exact instance-level pairing cannot be assumed. To handle this, the authors develop a dual-kernel contrastive learning framework that begins by reducing the distribution discrepancy across views on the aligned subset through similarity-driven alignment objectives. Drawing on the semantic priors derived from these aligned pairs, the approach then builds a cross-view kernel alignment guidance graph spanning both aligned and unaligned samples. Guided by this graph, a kernel-based contrastive objective encourages category-level semantic agreement without resorting to explicit instance matching. In parallel, a contrastive objective is also imposed at the cluster level. Lastly, the graph is used to fuse features across views, and the resulting representations are fed into k-means for clustering.

Critical strengths
  1. The paper targets a practically relevant and well-motivated problem, as partially aligned views frequently arise in real-world multi-modal and multi-sensor data.

  2. The framework sidesteps explicit one-to-one correspondence recovery, which tends to be fragile when only partial alignment is available.

  3. The empirical evaluation is fairly comprehensive, covering eight benchmark datasets under partial, full, and extremely low alignment ratios, together with ablation studies that examine the contribution of each loss component.

Critical weaknesses
  1. The paper does not specify how misalignment is constructed; presumably, following prior work, it is simulated by randomly shuffling samples, which may not adequately reflect the more structured or realistic misalignment patterns that arise in real applications.

  2. The paper asserts that the method carries out alignment in a Hilbert space through kernel contrastive learning. The authors should articulate more explicitly what benefits alignment in a Hilbert space offers over alignment in Euclidean space, and in particular why kernel contrastive learning in a Hilbert space leads to improved performance.

  3. The method is formulated only for the two-view case, and its extension to more views is described as straightforward without further details. Since several benchmarks contain more than two views, the authors should clarify how the alignment graph and the fused representation are constructed in the multi-view setting.

Questions for authors
  1. Could the authors add experiments under more structured misalignment settings, such as class-dependent or local permutations, and clarify how the unaligned data are constructed?

  2. Could the authors analyze why alignment in a Hilbert space outperforms alignment in Euclidean space, especially since, with ℓ2-normalized features and a bandwidth of 1, the Gaussian kernel is a monotonic function of cosine similarity?

  3. Could the authors explain how DKLPVC extends to more than two views, for example whether alignment graphs are built for every view pair or around an anchor view, and how sensitive the results are to this choice?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.