Abstract
This paper presents a new feature selection (FS) approach, called constructive FS approach (CFSA), based on the wrapper approach. The vital aspect of the new approach is that it determines both a set of features and neural network architecture for the features simultaneously. CFSA first divides the original feature set representing a problem into two clusters based on the correlation information. It puts similar features in one cluster, while dissimilar ones in the other cluster. The proposed algorithm then uses a constructive approach to find a subset of features from the two clusters and hidden neurons of the neural network. Five benchmark classification problems are used to evaluate the efficacy of CFSA. The experimental results exhibit that CFSA has ability to find salient features with a small number of hidden neurons to produce robust neural classifiers.
