Multiple k-NN Classifiers Fusion Based on Evidence Theory

Multiple classifiers fusion is a powerful solution to the difficult and complex classification problems, which can improve performance and generalization capability. This paper presents a multiple k-nearest neighbor classifiers fusion approach based on evidence theory. Independent k-NN classifiers are established based on heterogeneous features. The novel approach to generating mass functions of a given sample for each member classifiers are based on the class distributions on the k- nearest neighbors over heterogeneous features. Based on Dempster rule of combination, we can obtain the combined mass functions. Then the corresponding belief functions can be derived and the classification decisions of the fused classifier can be done. The approach proposed is promising because it takes full advantage of the simplicity of k-NN classifier and the better performance based on classifiers fusion. Experimental results provided show the efficacy and rationality of the approach proposed.