Feature Selection Based On Feature Fuzzify And Neural Networks

Feature selection is a key problem in pattern recognition system. Recently, it is popular to use artificial neural network to resolve such a problem. Most researchers think that a neural network learned well will give hints to the importance degree of a feature. Some proposed feature importance evaluations are network weights'functions. Some are evaluations of weights'saliency or input nodes'sensitivity. Others are based on both neural network technology and fuzzy set theory. Two feature selection algorithms based on neural network and fuzzy membership function are proposed to resolve some problems of others. Both algorithms map data from origin feature space to fuzzy feature space through auto adaptive fuzzy membership functions, where the classify problem will be easier and the data are normalized simultaneously. The first one is based on a fuzzy-neuro network with an auto adaptive fuzzify layer. The network is a fuzzy inference system with learning ability. The second one is a multilayer perceptron based on auto adaptive fuzzy membership functions. The method is similar with the above one but gets over the one's problem of processing data with high dimensions. The merits of above two methods are: 1) The fuzzy membership functions are auto adaptive. 2) The feature evaluation function is simple and easy to understand. 3) The network can be used to classify patterns and its actions are easy to analysis. 4) The seletion algorithm can form a feature selection system with feature space searching algorithms. The two methods are proved to be effective by theory analysis and experiments results.