Building Diverse Classifier Outputs to Evaluate the Behavior of Combination Methods: The Case of Two Classifiers

In this paper, we report an experimental comparison between two widely used combination methods, i.e. sum and product rules, in order to determine the relationship between their performance and classifier diversity. We focus on the behaviour of the considered combination rules for ensembles of classifiers with different performance and level of correlation. To this end, a simulation method is proposed to generate with fixed accuracies and diversity a set of two classifiers providing measurement outputs. A measure of distance is used to estimate the correlation among the pairs of classifiers. Our experimental results tend to show that with diverse classifiers providing no more than three solutions, the product rule outperforms the sum rule, whereas when classifiers provide more solutions, the sum rule becomes more interesting.