Classifier Selection Method Based on Multiple Diversity Measures

Abstract

Multiple Classifier Systems can make up for some defects of a single classifier. It has been widely used in machine learning, pattern recognition, and other fields. However, it is easy to generate some redundant classifiers with small difference, when the number of classifiers increases. In order to select classifiers with great diversity, a classifier selection method based on multiple diversity measures is proposed in this paper. Firstly, the fusion matrix is constructed by using five pairwise diversity measures. Then, the graph obtained by the fusion matrix is colored by the ant colony algorithm, and the candidate ensembles are generated. Finally, we introduce the fuzzy information theory and combine with five non-pairwise diversity measures to select a group of classifiers. The experimental results show that the proposed method is feasible and can significantly improve the accuracy of the ensemble.