Contrastive Subspace Clustering with Dissimilarity Regularization

Yu-Jie Chen, Wenhui Wu, Le Ou-Yang

Abstract: Subspace Clustering (SC) aims at clustering the data according to its underlying subspaces. However, existing subspace clustering methods fail to make full use of available information. Inspired by advances that contrastive learning has achieved in unsupervised representation learning, we propose a method, namely contrastive subspace clustering with dissimilarity regularization (CSCDR), incorporating contrastive learning into subspace clustering. Specifically, we first construct another view of data through augmentation. Then, the deep subspace clustering networks (DSC-Nets) are used to learn the self-expressive coefficients of data and its augmentation, which further interact with each other by contrastive learning. Moreover, a dissimilarity matrix is formulated based on the latent feature of data to further calibrate the noisy connections in the self-expressive coefficient. And we also utilize the relevant similarity matrix contributing to the multi-positive contrasitve learning, which ameliorates the cluster-collapse problem of representation. Experimental results on several data sets demonstrate the efficacy of the proposed method.