Specific-to-general Approach for Rule Induction Using Discernibility Based Dissimilarity

In this study, we propose a new decision rule induction approach. Conventional rule induction methods are often based on sequential covering with the general-to-specific approach in which to generate a premise of a rule, the premise is initialized to be empty and conditions are added to it until no or few negative objects are covered by the premise. While, in this study, we propose a rule induction method using the specific-to-general approach by applying discernibility based clustering to positive objects. In our approach, positive objects are clustered using a similarity measure which is related to discernibility of clusters. From an obtained cluster, we can generate a premise of a decision rule by taking common condition values of objects in the cluster.