Bidirectional fusion for deep contrastive multi-view clustering

Abstract

In recent years, contrastive learning has found applications in multi-view clustering. Although these methods have achieved some performance improvements, they still suffer from the negative impact of incorrect contrastive pairs. Similar to many traditional multi-view clustering methods that focus solely on either similarity matrices or feature matrices, existing contrastive learning methods often emphasize learning from the perspective of feature matrices. This unidirectional approach limits the selection of high-quality contrastive samples. To address these challenges, we propose a novel bidirectional fusional deep contrastive multi-view clustering method (BFCMC). Specifically, BFCMC simultaneously focuses on similarity matrices and low-dimensional feature matrices to learn a clearer, ground truth-aligned unified affinity matrix. Employing this matrix to guide the selection of contrastive samples effectively addresses the issue of incorrect contrastive pairs. Building on this, we propose a bidirectional fusion contrastive learning strategy that incorporates intra-view modules to enhance feature discrimination and inter-view modules to ensure representation consistency. Extensive experiments on multiple real-world datasets demonstrate the superiority of BFCMC compared to state-of-the-art methods.