Pair-wise Diversity Measure for Multiple Classifier System Based on Evidence Theory

A method of designing pattern recognition systems, known as the multiple classifier system, has emerged over recent years to address the practical problem of designing classification systems with improved accuracy and efficiency. The aim is to design a composite system that outperforms any individual classifier by pooling together the decisions of all classifiers. Diversity among base classifiers is known to be a necessary condition for improvement in multiple classifier system ensemble performance. In this paper, a new method of diversity measure for multiple classifier system is proposed and a new approach to multiple classifier system design is presented. For a given test sample, the diversity measure for each pair of sub-classifier is constructed based on the distance of basic probability assignment. The average value of all pair-wise classifiers is the final result of diversity measure. We combine the proposed new method of diversity measure with the traditional diversity measure, constructing an effective multiple classifier system. Experiments on UCI data sets and SAR images data sets show that the proposed diversity measure and the proposed approach to construct multiple classifier system are rational and effective.