Feature Selection Based on Sparse Fisher Discrimimant Analysis

This paper proposes a novel method of sparse Fisher linear discriminant analysis (SFLDA) for dimensionality reduction. Utilizing the equivalence of Fisher linear discriminant analysis (FLDA) and least squares linear regression (LSLR), sparse Fisher linear discriminant vector can be obtained by introducing L₁ regularization into a least squares error criterion function. The sparse Fisher linear discriminant vector has only a small number of nonzero components. This implies that the sparse discriminant vector learned by SFLDA has a more intuitionistic physical interpretation than the dense one. The feasibility and effectiveness of the proposed method is verified on 3 real-world data sets from UCI, USPS handwriting digital data set and AR face database with competative or better results.