Deep dual contrastive learning for multi-view subspace clustering

Abstract

Multi-view subspace clustering (MVSC) aims to learn a consistent shared self-representation by utilizing the consistency and complementarity of all views, numerous MVSC algorithms have attempted to obtain the optimal representation directly from raw features. However, they might overlook the noisy or redundant information in raw feature space, resulting in learning suboptimal self-representation and poor performance. To address this limitation, an intuitive idea is introducing deep neural networks to eliminate the noise and redundancy, yielding a potential embedding space. Nevertheless, existing deep MVSC methods merely focus on either the embeddings or self-expressions to explore the complementary information, which hinders subspace learning. In this paper, we present a deep multi-view dual contrastive subspace clustering framework to exploit the complementarity to learn latent self-representations effectively. Specifically, multi-view encoders are constructed to eliminate noise and redundancy of the original features and capture low-dimensional subspace embeddings, from which the self-representations are learned. Moreover, two diverse specific fusion methods are conducted on the latent subspace embeddings and the self-expressions to learn shared self-representations, and dual contrastive constraints are proposed to fully exploit the complementarity among views. Extensive experiments are conducted to verify the effectiveness of the proposed method.