In this paper, we present an empirical analysis on the relationship between diversity and accuracy of classifier ensembles in the context of the theory of belief functions. We provide a modelling for formulating classifier outputs as triplet mass functions and a unified notation for defining diversity measures and then assess the correlation between the diversity obtained by four pairwise and non-pairwise diversity measures and the improvement of accuracy of classifiers combined in decreasing and mixed orders by Dempster’s rule, Proportion and Yager’s rules. Our experimental results reveal that the improved accuracy of classifiers combined by Dempster’s rule is positively correlated with the diversity obtained by the four measures, but the correlation between the diversity and the improved accuracy of the ensembles constructed by Proportion and Yager’s rules is negative, which is not in favor of the claim that increasing diversity could lead to reduction of generalization error of classifier ensembles.
|Title of host publication||Multiple Classifier Systems Lecture Notes in Computer Science|
|Publication status||Published - 2011|