Redundancy-Constrained feature selection with radial basis function networks

Neural Networks are widely used to select features for classification / regression problems. These methods usually do not take into account the redundancy (linear/nonlinear dependency) between features. Consequently the selected set of features although useful, may contain redundant features. Here we propose a general framework for feature selection with controlled redundancy using a radial basis function (RBF) network. We demonstrate the effectiveness of the method on some benchmark data sets. Our framework can be easily adapted to other neural networks.