DWCL: Dual-Weighted Contrastive Learning for Multi-View Clustering

TL;DR

DWCL addresses representation degeneration in multi-view clustering by introducing a Best-Other (B-O) contrastive mechanism and a dual weighting scheme that jointly account for view quality and view discrepancy. By selecting the best view via silhouette coefficient and pairing it with other views, DWCL reduces cross-view complexity from $O(|V|^2)$ to $O(|V|)$ while down-weighting unreliable cross-views. The dual weight combines the view quality weight $W_{SI}$ and the discrepancy weight $W_{CMI}$ to strengthen high-quality, low-discrepancy cross-views, with theoretical proofs linking the framework to mutual information optimization. Extensive experiments across eight datasets show DWCL consistently outperforms state-of-the-art MVCC methods, including notable improvements on Caltech5V7 and CIFAR10, and ablations confirm the effectiveness of both components. The approach offers a scalable, robust path for multi-view clustering with improved representation learning and clustering accuracy.

Abstract

Multi-view contrastive clustering (MVCC) has gained significant attention for generating consistent clustering structures from multiple views through contrastive learning. However, most existing MVCC methods create cross-views by combining any two views, leading to a high volume of unreliable pairs. Furthermore, these approaches often overlook discrepancies in multi-view representations, resulting in representation degeneration. To address these challenges, we introduce a novel model called Dual-Weighted Contrastive Learning (DWCL) for Multi-View Clustering. Specifically, to reduce the impact of unreliable cross-views, we introduce an innovative Best-Other (B-O) contrastive mechanism that enhances the representation of individual views at a low computational cost. Furthermore, we develop a dual weighting strategy that combines a view quality weight, reflecting the quality of each view, with a view discrepancy weight. This approach effectively mitigates representation degeneration by downplaying cross-views that are both low in quality and high in discrepancy. We theoretically validate the efficiency of the B-O contrastive mechanism and the effectiveness of the dual weighting strategy. Extensive experiments demonstrate that DWCL outperforms previous methods across eight multi-view datasets, showcasing superior performance and robustness in MVCC. Specifically, our method achieves absolute accuracy improvements of 5.4\% and 5.6\% compared to state-of-the-art methods on the Caltech6V7 and MSRCv1 datasets, respectively.

Figures (9)

• Figure 1: Comparison of our DWCL with existing works in cross-view construction. (a) Pairwise cross-view construction often results in a large number of unreliable cross-views, which limits the representation of individual views. (b) Self-weighted cross-view construction, based on view discrepancy, can unintentionally amplify low-quality cross-views that exhibit low discrepancy in contrastive learning. (c) In contrast, our approach employs the silhouette coefficient to create dual-weighted cross-views, ensuring that the representation learning of individual views is guided by the highest-quality view.
• Figure 2: The framework of DWCL. Silhouette coefficient (SI), the internal evaluation index of k-means clustering, is utilized to determine the best view. In our Best-Other (B-O) contrastive mechanism, specific cross-views are formed by combining the best view with other views. The view quality weight, $\mathcal{W}{SI}$, is integrated with the view discrepancy weight, $\mathcal{W}{CMI}$, to effectively reduce the influence of low-quality and high-discrepancy cross-views in the dual-weighted contrastive learning.
• Figure 3: Comparison of DWCL with seven baseline methods and BSV across eight datasets. The vertical axis represents the difference in ACC between each baseline and the BSV for eight multi-view datasets.
• Figure 4: (a) Initial accuracy of each view on Caltech6V7. (b) Changes in weights $\mathcal{W}^{1,2}$, $\mathcal{W}^{1,4}$, $\mathcal{W}^{2,4}$ and $\mathcal{W}^{4,5}$ in the traditional pairwise and B-O contrastive mechanisms. (c) Clustering performance (ACC) of views 1, 2, 4, and 5 in the traditional pairwise and B-O contrastive mechanisms.
• Figure 5: Convergence Results of the Best View Compared to Each View Obtained by DWCL on Caltech6V7.