Sparse kernel fisher discriminant analysis

Abstract

This paper presents a method of Sparse Kernel Fisher Discriminant Analysis (SKFDA) through approximating the implicit within-class scatter matrix in feature space. Traditional Kernel Fisher Discriminant Analysis (KFDA) has to use all the training samples to construct the implicit within-class scatter matrix while SKFDA needs only small part of them. Based on this idea, the aim of sparseness can be obtained. Experiments show that SKFDA can dramatically reduce the number of training samples used for constructing the implicit within-class scatter matrix. Numerical simulations on "Banana Shaped" and "Ripley and Ionosphere" data sets confirm that SKFDA has the merit of decreasing the training complexity of KFDA. © Springer-Verlag Berlin Heidelberg 2005.