Low-Rank Attention and Contrastive Alignment for Deep Multi-View Clustering

ICLR 2026 Conference Submission18061 Authors

Keywords: Multi-view clustering, Low-rank attention, Contrastive learning, Deep clustering

Abstract: Recent years have witnessed significant advancements in deep multi-view clustering (MVC). However, prevailing methods exhibit three critical limitations: (1) poor scalability for large-scale datasets, (2) neglect of anchor semantic consistency in feature alignment, and (3) inability to capture high-order feature interactions. To overcome these challenges, we propose a Low-Rank Attention and Contrastive Alignment framework (LRACA). Unlike conventional approaches that align sample-level features in shared subspaces, LRACA employs a category-aware anchor generation module to directly align high-level semantic prototypes (i.e., category centers) across views, explicitly enforcing clustering semantic consistency. Furthermore, we devise a dynamic low-rank attention mechanism to enhance feature discriminability, where entropy regularization constrains attention weight distributions to derive clustering pseudo-labels. Finally, a pseudo-label-guided cluster-level contrastive learning module maximizes cross-view mutual information through a feed-forward optimization paradigm. Extensive experiments on six large-scale multi-view datasets demonstrate that LRACA significantly outperforms state-of-the-art methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 18061