On The Use of Nonparametric Neighborhood Classification Rules in Multiple Classifier Combination

Abstract

A multiple classifier combination approach based on nonparametric neighborhood classifiers is proposed in this paper. Two different types of nonparametric neighborhood classifiers are used for each query sample, which can be regarded as two different sources of evidence. One type of member classifier emphasizes the similarity and the other type emphasizes the spatial distribution in training set with respect to the query sample. Two mass functions then can be determined based on two different mass function generation methods proposed. According to evidence combination, better classification accuracy can be obtained. The approach proposed has no problem of parameter optimization or selection. In the experiments, the efficacy and rationality of the methods proposed are verified.