Multivariate Pattern Classification based on Local Discriminant Component Analysis

This paper proposes a novel local discriminant component analysis (DCA) algorithm that is useful for pattern classification of high-dimensional data. Different from most traditional methods, in which feature extractors are usually used prior to a classifier, the proposed method incorporates the feature extraction process into the classifier. Then, a probabilistic neural network is developed based on the idea of local DCA, in which the whole network including the feature extractor and the classifier can be modulated according to a single training criterion, so that features suited to the classification purpose can be extracted. In this paper, a hybrid training algorithm is proposed on the basis of the minimum classification error (MCE) learning. In simulation experiments, benchmark data are used to prove feasibility of the proposed method