Determining the saliency of input variables in neural network classifiers

This paper examines a measure of the saliency of the input variables that is based upon the connection weights of the neural network. Using Monte Carlo simulation techniques, a comparison of this method with the traditional stepwise variable selection rule in Fisher's linear classification analysis (FLDA) is made. It is found that the method works quite well in identifying significant variables under a variety of experimental conditions, including neural network architectures and data configurations. In addition, data from acquired and liquidated firms is used to illustrate and validate the technique.