Coevolution for classification

Abstract

A data mining field with daily, and sometimes even vital, practical applications, classification has been addressed by many powerful paradigms, among which evolutionary algorithms (EAs) play a successful role. Nevertheless, as evolutionary computation (EC) progresses, there appear new possibilities of developing simpler and yet robust classification techniques. The aim of this paper is hence to put forward a novel evolutionary classification framework which embodies two contradictory prototypes coming from the state-of-the-art field of coevolution and which has proven to be a viable alternative. Coevolution between individuals assumes two opposite interactions: cooperative and competitive. Analogously, coevolution for classification assumes two possible and opposed manners of solving the task. Within both approaches, the solution of a classification problem is regarded as a set of IFTHEN conjunctive rules in first order logic. As a consequence, learning will be driven either by the cooperation between rules towards a complete and accurate rule set or by the competition between rules and training samples in the direction of extensive and hard testing on each side. The paper is organized as follows. The next section introduces a general point of view upon classification. Section three brings an overview on coevolution: The cooperative and competitive archetypes are outlined and explained. Section four describes the proposed manner of approaching classification from the cooperative side, while section five presents the application of the competitive counterpart. Experiments on three data sets, two benchmark and one real-world, are depicted in section six and the paper closes with the concluding remarks.