Multiple Classifiers Fusion Based on Weighted Evidence Combination

Abstract

Multiple Classifiers Fusion is to utilize distinguished classifiers to resolve the same classification problem as a single classifier does, which can improve performance and generalization capability. In this paper, a new method of multiple classifiers fusion based on weighted evidence combination is proposed. Independent member classifiers are designed based on heterogeneous features by utilizing Artificial Neural Network (ANN). The Basic Probability Assignments (BPA or mass function) are generated based on member classifiers' outputs corresponding to a given test sample. The weights of each member classifier are defined based on their respective class-wise classification performance on training dataset. Based on weighted evidence combination, classification results of the fused classifier can be obtained, which is better than those derived based on Dempster rule of combination without weights. The experimental results provided in this paper verify the rationality and efficacy of the method proposed.