A new method for feature selection

Abstract

We present a new approach based on discriminant analysis and regularization neural network for salient feature selection. Using the discriminant analysis based feature ranking, an ordered feature queue can be obtained according to the saliency of features. The neural network is trained by minimizing an augmented cross-entropy error function in the method. Feature selection is based on the reaction of the cross-validation data set classification error due to the removal of the individual features. The approach proposed is compared with four other feature selection methods, each of which banks on a different concept. The algorithm proposed outperforms the other methods by achieving higher classification accuracy on all the problems tested. © Springer-Verlag Berlin Heidelberg 2006.