Learning latent disentangled embeddings and graphs for multi-view clustering

Abstract

Graph based methods have recently attracted much attention for multi-view clustering. Most existing methods seek the latent shared embeddings to learn a unified similarity graph or fuse multiple view-specific graphs to a consensus one for clustering, which may not sufficiently explore the common and complementary information among views. Besides, the high-order inter-view correlations are not fully investigated. To address these issues, this paper proposes a latent Disentangled Embeddings and GRaphs based multi-viEw clustEring (DEGREE) method, which considers the common and view-specific information in a latent subspace by explicit embedding disentanglement and multiple graphs learning. We assume that each view can be generated from a shared latent embedding and a corresponding view-specific embedding, which model the common information and exclusive complementary information among views, respectively. The intra-view and inter-view exclusivities among embeddings are encouraged by an orthogonality regularizer. To fully use the underlying information, we excavate the pairwise instance relations in both shared embedding and diverse view-specific embeddings by learning multiple graphs. Besides, a tensor singular value decomposition (t-SVD) based tensor nuclear norm regularizer is imposed on view-specific graphs, which helps to explore the high-order inter-view correlations. An alternative optimization algorithm is designed to solve the proposed model. Experimental evaluations on several popular datasets demonstrate that our DEGREE method outperforms the state-of-the-art methods.