Contrastive Alignment for Deep Multi-view Subspace Clustering

Deep multi-view subspace clustering can partition non-linear high-dimensional data by exploring a consensus self-expression among the samples and has attracted increasing attention. However, most existing methods learn the consensus self-expression by fusing view-specific self-expression, where they neglect that view-private meaningless information or noise may interfere the learning of consensus self-expression. In this paper, we propose a novel Contrastive Alignment for Deep Multi-view Subspace Clustering (CADMVSC) method. Specifically, the view-specific self-expression are obtained by self-expressive layer embedded in autoencoder network for each view. To guarantee the consistency among the views and reduce the interference of view-private information, we align the view-specific self-expression by contrastive learning. Moreover, all the aligned self-expression are fused adaptively to obtain the consensus self-expression by channel attention mechanism, which can fully explore the complementarity of views. Extensive experiments have shown the excellent clustering performance of our method over other state-of-the-art methods.