Combination of heterogeneous multiple classifiers based on evidence theory

In the field of multiple classifiers combination, diversity among member classifiers is known to be a necessary condition for improving ensemble performance. In this paper we use different types of member classifiers based on heterogeneous features to increase the diversity when we implement the multiple classifier system (MCS). Member classifiers adopted in this paper include the k-NN classifier and the BP network classifier. The combination algorithm is based on Dempster rule of combination. The approaches to generating mass functions corresponding to the type of member classifiers are proposed. It is shown experimentally that the proposed approaches are rational and effective. The approaches proposed in this paper provide a new way to combine the two different types of classifiers: the k-NN classifiers and the BP network classifiers. Thus their corresponding strengths can be fully utilized and their corresponding drawbacks can be counteracted.