Deep Contrastive Multi-view Subspace Clustering

Abstract

Multi-view subspace clustering has become a hot unsupervised learning task, since it could fuse complementary multi-view information from multiple data effectively. However, most existing methods either fail to incorporate the clustering process into the feature learning process, or cannot integrate multi-view relationships well into the data reconstruction process, which thus damages the final clustering performance. To overcome the above shortcomings, we propose the deep contrastive multi-view subspace clustering method (DCMSC), which is the first attempt to integrate the contrastive learning into deep multi-view subspace clustering. Specifically, DCMSC includes multiple autoencoders for self-expression learning to learn self-representation matrices for multiple views which would be fused into one unified self-representation matrix to effectively utilize the consistency and complementarity of multiple views. Meanwhile, to further exploit multi-view relations, DCMSC also introduces contrastive learning into multi-autoencoder network and Hilbert Schmidt Independence Criterion (HSIC) to better exploit complementarity. Extensive experiments on several real-world multi-view datasets demonstrate the effectiveness of our proposed method by comparing with state-of-the-art multi-view clustering methods.