Bare bones particle swarm optimization with adaptive chaotic jump for feature selection in classification

Feature selection (FS) is a crucial data pre-processing process in classification problems. It aims to reduce the dimensionality of the problem by eliminating irrelevant or redundant features while achieve similar or even higher classification accuracy than using all the features. As a variant of particle swarm optimization (PSO), Bare bones particle swarm optimization (BBPSO) is a simple but very powerful optimizer. However, it also suffers from premature convergence like other PSO algorithms, especially in high-dimensional optimization problems. In order to improve its performance in FS problems, this paper proposes a novel BBPSO based FS method called BBPSO-ACJ. An adaptive chaotic jump strategy is designed to help the stagnated particles make a large change in their searching trajectory. It can enrich the search behavior of BBPSO and prevent the particles from being trapped into local attractors. A new global best updating mechanism is employed to reduce the size of obtained feature subset. The proposed BBPSO-ACJ is compared with eight evolutionary computation (EC) based wrapper methods and two filter methods on nine benchmark datasets with different number of dimensions and instances. The experimental results indicate that the proposed method can select the most discriminative features from the entire feature set and achieve significantly better classification performance than other comparative methods.