Evolutionary Feature Selection Based on Semi-Local Search

摘要

Feature selection by extracting the most informative features from a dataset improves the accuracy of a classifier, reduces its complexity and helps to speed up the classification tasks. In this paper, a new hybrid feature selection method based on a combination of genetic algorithm (GA) and particle swarm optimization (PSO) is introduced. In this structure, instead of point to point search used in most of the search methods, the subspaces are sequentially determined by an enhanced genetic algorithm, where each subspace is efficiently searched by a PSO method. In the proposed GA, each chromosome is equal to a subspace of the search space. Crossover and mutation operators over the defined chromosomes generate new subspaces. PSO algorithm searches in the zone and returns the fitness value of the corresponding chromosome. The idea of defining subspaces is very efficient in utilizing the exploration ability of GA. In addition, PSO exploitation reduces the time complexity of the pure genetic search. Reported results on 10 UCI benchmark datasets confirm how this method has significant improvement in classification performance.