Unified Multi-view Clustering based on Joint Multi-Structure Representation Learning

Recently, tensor-based multi-view subspace clustering methods have shown promising results. Despite the impressive clustering performance, most clustering methods still face the following challenges: 1) the expensive computational burden caused by complex clustering steps, 2) the limitation of scalability to large-scale data sets, and 3) the failure to fully explore inter/intra-view information by tensor rank approximations. To address these issues, we propose a joint multi-structure representation learning for multi-view clustering (JMSR-MVC), where the low-dimensional latent representation structure for each view is constructed by a matrix factorization mechanism. The high-order correlated inter/intra-view structure information is explored by the low-rank tensor multi-scale entanglement renormalization ansatz (MERA). In addition, the K-means clustering structure is introduced into the model, which integrates the high-order tensor representation learning and the corresponding latent cluster assignments into a unified manipulation instead of traditional independent two-stage scheme. The alternating direction method of multipliers algorithm is developed to solve the optimization model. Furthermore, we give the analysis of the computational complexity and convergence. Extensive experiments on six challenging data sets demonstrate the superiority of the proposed method over current state-of-the-art multi-view clustering approaches.