Research On The Combining Methods And Diversity Measures In Multiple Classifier Systems

Since multiple classifier systems can to some extent improve the performance of classification, the technique has been widely used in various fields of pattern recognition. Researchers have propsosed a number of combining methods, so it's necessary to make a survey on them. In addition, the recent research on multiple classifier systems is no longer limited to proposing or enhancing the combining methods. More and more attention is paid for the relationship among base classifiers, especially the diversity in classifier ensembles. Researchers want to find a proper diversity measure, whose ability in predicting the performance of classifer combining can be used to assist the construction of multiple classifier systems.This paper studies these two main problems in multiple classifier systems: combining methods and diversity measures. The main contributions of the work are as follows:Firstly, a survey on the combining methods in multiple classifier systems is given, in which the advantages and disadvantages of each method are summarized. In addition, we design some experiments to compare the methods.Secondly, a new kind of diversity measure for classifier ensembles named EPD is proposed. The experiments on UCI Machine Learning Repository prove that, compared to existing measures, EPD shows stronger ability in predicting the performance of multiple classifier systems.Thirdly, an algorithm of ensemble thinning based on EPD is developed and used to construct an improved face recognition system. The experiments on AT&T face database prove that, using the algorithm, though the number of base classifiers in the classifier ensemble is reduced, the error rate can be maintained or even decreased.