On the utility of imprecise rules induced by MLEM2 in classification

Rules inferring the memberships to single decision classes have been induced in rough set approaches and used to build a classifier system. Rules inferring the memberships to unions of multiple decision classes can be also induced in the same manner. In this paper, we show the classifier system with rules about the union of multiple decision classes has an advantage in the accuracy of classification. However, those rules are not always practical because the number of those rules becomes much more than that of rules inferring the memberships to single decision classes. We examine several methods for reducing of the number of rules about the union of multiple decision classes and discuss whether the classification accuracy is preserved or not. By numerical experiments, we investigate also what factor is related to the classification accuracy. To this end, we consider the distance from the class distribution, the robustness of the classification and the similarity between combined classes as factors.