A Forward Search Inspired Particle Swarm Optimization Algorithm for Feature Selection in Classification

Particle swarm optimization (PSO) has been widely used for feature selection (FS) in classification. However, FS is still a challenging optimization task for PSO when the dimensionality of data is high. In this paper, we propose a forward search inspired PSO (FSIPSO) algorithm to build a wrapper-based FS method. In FSIPSO, the search space dynamically changes during the evolutionary process. Specifically, we rank the features according to their single-feature classification performance and divide the search space into several sub-spaces. A forward search scheme is proposed to sequentially select the sub-spaces. The selected sub-spaces construct the search space for FSIPSO. With this scheme, FSIPSO first searches in a small space to quickly find candidate solutions (feature subsets) with relatively good performance. Then, the search space expands with the selection of more sub-spaces, and FSIPSO can further select informative features in the expanded search space. Moreover, mutation operations are used in FSIPSO to avoid the premature problem. The experimental results on 8 UCI datasets have shown that FSIPSO obtains better FS results with less computation time compared with benchmark PSO-based FS methods. FSIPSO also obtains better convergence performance than these methods.