Feature selection: a neural approach

Feature selection is an integral part of most learning algorithms. By selecting relevant features of the data, higher predictive accuracy or classification rate can be expected from a machine learning method. We propose an approach to feature selection based on neural network pruning. The method performs a backward selection by successively removing input nodes in a network trained with the complete set of features as inputs. When an input node is removed, and relative weight connections are excised, the remaining weights are updated so as to keep approximately unchanged the behavior of the network. A simple criterion to select input nodes to be removed is developed. Experimental results over a well-known classification problem show the feasibility of the proposed approach and encourage its application to other classification tasks