Adaptive Weight Structure Representation for Multi-view Subspace Clustering

Abstract

Traditional subspace clustering method based on self-representation has been widely applied in learning similarity matrix. Existing self-representation based methods treat all features equally in the process of learning similarity matrix, which makes redundant features in learning stage may have a certain negative impact on the final representation and even the representation of other non-redundant features. To solve the above problems, this paper proposes Adaptive Weight Low-Rank Representation (AWSLRR) algorithm. AWSLRR uses firstly nested structure to learn more clean and reasonable similarity matrix, then applied weight constraints to the reconstruction corruption which is generated during the reconstruction process. The adaptive weight matrix imposes a small weight coefficient on the larger corruption value by imposing constraints on the corruption term during the self-representation process, and vice versa. Finally, the experimental results on five real datasets validate the competitiveness of proposed algorithm.