Section: New Results

Local Search in the Context of Classification Rule Mining

Participants: Julie Jacques, Laetitia Jourdan, Clarisse Dhaenens

Many multi-objective algorithms have been proposed to solve the classification rule mining problem; the vast majority of them are based on genetic algorithms. We propose an algorithm, MOCA - Multi-Objective Classification Algorithm -, to solve this problem. The originality of MOCA is to be a dominance-based multi-objective local search (DMLS) using a Pittsburgh representation of rules. We evaluated several DMLS implementations and neighborhood operators on literature datasets and one real dataset. Then we compared the best obtained algorithm against several efficient approaches of the literature. The experiments show that the proposed approach is very competitive in comparison to other algorithms tested. Moreover, our approach is able to deal with very large real datasets and manages to have a good accuracy.