Constrained Low-Rank Tensor Learning for Multi-View Subspace Clustering

This paper studies multi-view subspace clustering. Existing low-rank tensor learning methods employ t-SVD to explore the multi-view correlation information. However, the local structure in the view-specific feature spaces is ignored, and the self-representation coefficients and affinity matrix are often learned in two distinct processes. To solve the two problems, we propose constrained low-rank tensor learning (CLRTL). First, we adopt t-SVD to preserve the global low-rank structure and multiple views correlation. Meanwhile, a sparse congruency constraint using the l(2,1)-norm is imposed on the coefficient matrix of each view to reveal the intra-view local structure. Furthermore, the affinity matrix is joined to the optimization framework to directly learn the correlation between samples. Last, we develop an efficient algorithm to deal with the proposed problem. Experimental results on the benchmark datasets verify the superiority of CLRTL.