Feature subset selection using constructive neural nets with minimal computation by measuring contribution

Abstract

In this paper we propose a new approach to select feature subset based on contribution of input attributes in a three-layered feedforward neural network (NN). Three techniques: constructive, contribution, and backward elimination are integrated together in this method. Initially, to determine the minimal NN architecture, the number of hidden neurons is determined by a constructive approach. After that, one-by-one removal of input attributes is performed to compute their contribution. Finally, a sequential backward elimination is used to generate relevant feature subset from the original input space. The elimination process is continued depending on a criterion. To evaluate the proposed method, we applied it to four real-world benchmark problems. Experimental results confirmed that, the proposed method significantly reduces the irrelevant features without degrading the network performance and generates the feature subset with good generalization ability. © 2008 Springer-Verlag Berlin Heidelberg.