Dimensionality Reduction via Locality Constrained Competitive Sparse Representation by L2-Norm Regularization

Abstract

Dimensionality reduction is an important scheme for dealing with data of high dimensionality. Subspace learning is extensively studied in pattern recognition due to the fact that it is simple and computational efficient. In this paper, a new dimensionality reduction method is proposed based on sparse representation. First, the representation coefficients are obtained through competitive sparse representation with L2-Norm regularization. To make the coefficients more discriminant, the locality structure is also utilized. Then, the inter-class graph (or the penalty graph) and intra-class graph (or the intrinsic graph) are constructed according to the coefficients. Finally, the optimal projective matrix is calculated by minimizing the intar-class scatter while maximizing the inter-class scatter. The experimental results show that the proposed method exhibits good performances.