Rotational Linear Discriminant Analysis Using Bayes Rule for Dimensionality Reduction

Abstract

Linear discriminant analysis (LDA) finds an orientation that projects high dimensional feature vectors to reduced dimensional feature space in such a way that the overlapping between the classes in this feature space is minimum. This overlapping is usually finite and produces finite classification error which is further minimized by rotational LDA technique. This rotational LDA technique rotates the classes individually in the original feature space in a manner that enables further reduction of error. In this paper we present an extension of the rotational LDA technique by utilizing Bayes decision theory for class separation which improves the classification performance even further.