A New Diversity Measure for Classifier Fusion

The combination of multiple classifiers is one of the important topics in recent pattern recognition research. It is deemed that the diversity among the classifiers is a key issue in classifier combination. There are different diversity measures presented in the literature. In this paper, we first analyze several measures based on the oracle outputs of base classifiers while majority vote is taken as the combination method. Afterwards, we propose a new measure which overcomes the defects of the diversity measures analyzed. In order to illustrate the effectiveness of the proposed measure, a genetic algorithm based ensemble learning method is designed and compared on some UCI standard datasets to Bagging and Adaboost. The simulation shows that the proposed ensemble learning method gains better generalization performance over Bagging and Adaboost.