Efficient Direct Structured Subspace Clustering

Abstract

Subspace clustering splits data instances that are drawn from special low-dimensional subspaces via utilizing similarities between them. Traditional methods contain two steps: (1) learning the affinity matrix and (2) clustering on the affinity matrix. Although these two steps can alternatively contribute to each other, there exist heavy dependencies between the performance and the initial quality of affinity matrix. In this paper, we propose an efficient direct structured subspace clustering approach to reduce the quality effects of the affinity matrix on performances. We first analyze the connection between the affinity and partition matrices, and then fuse the computation of affinity and partition matrices. This fusion allows better preserving the subspace structures which help strengthen connections between data points in the same subspaces. In addition, we introduce an algorithm to optimize our proposed method. We conduct comparative experiments on multiple data sets with state-of-the-art methods. Our method achieves better or comparable performances.