Exploiting quadratic mutual information for discriminant analysis

Abstract

Novel criteria that reformulate the Quadratic Mutual Information according to Fisher's Discriminant Analysis are proposed for supervised dimensionality reduction. The proposed method uses a quadratic divergence measure and requires no prior assumptions about class densities. The criteria are optimized using gradient ascent with initialization using random or LDA based projections. Experiments on various datasets are conducted and highlight the superiority of the proposed approach compared to the standard QMI criterion. © 2012 Springer-Verlag.