A comparative study of neural-network feature weighting

AbstractMany feature weighting methods have been proposed to evaluate feature saliencies in recent years. Neural-network (NN) feature weighting, as a supervised method, is founded upon the mapping from input features to output decisions, and implemented by evaluating the sensitivity of network outputs to its inputs. Through training on sample data, NN implicitly embodies the saliencies of input features. The partial derivatives of the outputs with respect to the inputs in the trained NN are calculated to measure their sensitivities to input features, which means that implicit feature weighting of the NN is transformed into explicit feature weighting. The purpose of this paper is to further probe into the principle of NN feature weighting, and evaluate its performance through a comparative study between NN feature weighting method and state-of-art weighting methods in the same working conditions. The motivation of this study is inspired by the lack of direct and comprehensive comparison studies of NN feature weighting method. Experiments in UCI repository data sets, face data sets and self-built data sets show that NN feature weighting method achieves superior performance in different conditions and has promising prospects. Compared with the other existing methods, NN feature weighting method can be used in more complex conditions, provided that NN can work in those conditions. As decision data, output data can be labels, reals or integers. Especially, feature weights can be calculated without the discretization of outputs in the condition of continuous outputs.