Block Diagonal Sparse Subspace Clustering

Abstract

Sparse subspace clustering (SSC) is a spectral-type clustering-based method, which deals with high dimensional data via sparse representation. When the subspaces are independent of each other, the coefficient matrix obtained by SSC satisfies the block diagonal structure, which can better reveal the subspace attributes of data. In the actual environment, due to noise data and dependent subspaces, the obtained block diagonal structure is easy to be destroyed. To address the problem, we proposed the BDSSC method, which directly imposes the k-block diagonal regularizer on the coefficient matrix to purse the block diagonal structure. With the help of sparse prior and k-block diagonal regularizer, the coefficient matrix has a better block diagonal structure, and then the clustering performance is improved. Experiments on several actual datasets indicate that the proposed BDSSC method is superior to other state-of-the-art methods.