A fast anchor-based graph-regularized low-rank representation approach for large-scale subspace clustering

Abstract

Graph-regularized low-rank representation (GLRR) is an important subspace clustering (SC) algorithm, which has been widely used in pattern recognition and other related fields. It can not only represent the global structure of data, but also capture the nonlinear geometric information. However, GLRR has encountered bottlenecks in dealing with large-scale SC problems since it contains singular value decomposition and similarity matrix construction. To solve this problem, we propose a novel method, i.e., fast anchor-based graph-regularized low-rank representation (FA-GLRR) approach for large-scale subspace clustering. Specifically, anchor graph is first used to accelerate the construction of similarity matrix, and then, some equivalent transformations are given to transform large-scale problems into small-scale problems. These two strategies reduce the computational complexity of GLRR dramatically. Experiments on several common datasets demonstrate the superiority of FA-GLRR in terms of time performance and clustering performance.