Principal Component Analysis for Supervised Learning: a minimum classification error approach

Abstract

We present an alternative method to use Principal Component Analysis (PCA) for supervised learning. The proposed method extract features similarly to PCA but the features are selected by minimizing the Bayes error rate for classification. We show that the proposed method selects features that best separate the elements of the different classes. Using real and synthetic datasets, along with four different classifiers, experimental results show that the recognition accuracy of the proposed technique is improved compared to PCA.