Evolutionary Approaches for the Extraction of Classification Rules

Abstract

This paper provides evolutionary approaches in order to extract comprehensible and accurate classification rules. Indeed to construct a model of classification tone must extract not only accurate rules but comprehensible also, to help the human interpretation of the model and the decision make process. In this paper the authors describe a purely genetic approach, then a tabu search approach and finaly a memetic algorithm to extract classification rules. The memetic approach is a hybridization of a genetic algorithm (GA) and a local search based on a tabu search algorithm. Knowing that the amount of treated data is always huge in data mining applications, the authors propose to decrease the running time of the GA using a parallel scheme. In the authors' scheme the concept of generation has been removed and replaced by the cycle one and each individual owns a lifespan represented by a number of cycles affected to it randomly at its birth and at the end of which it disappears from the population. Consequently, only certain individuals of the population are evaluated within each iteration of the algorithm and not all our heterogeneous population. This causes the substantial reduction of the total running time of the algorithm since the evaluations of all individuals of each generation necessitates more than 80% of the total running time of a classical GA. This approach has been developed with the goal to present a new and efficient parallel scheme of the classical GA with better performances in terms of running time.