Sparse discriminative least squares regression model

In this paper, we propose a sparse discriminative least squares regression (SDLSR) for multicategory classification. The main mechanism of SDLSR is to apply L1-norm for regularizing the objective function of discriminative least squares regression (DLSR). Through theoretical analysis, it is obtained that SDLSR can be viewed as a relaxation of L1-SVM. Moreover, we theoretically prove that sparsity of objective function can effectively decrease the number of support vectors (SV). To optimize our proposed SDLSR, an efficient iterative algorithm is proposed. The experimental results confirm the effectiveness of our proposed SDLSR.