A dividing-based many-objective evolutionary algorithm for large-scale feature selection

Feature selection is a critical preprocess for constructing model in computer vision and machine learning, yet it is difficult to simultaneously satisfy both reducing features' number and maintaining classification accuracy. Toward this problem, we propose dividing-based many-objective evolutionary algorithm for large-scale feature selection (DMEA-FS). Firstly, four novel objectives are established for exploring the optimal feature's subsets. Meanwhile, we design two structures of wrapper for high accuracy and filter for low computation cost in DMEA-FS. Secondly, two new recombination methods are presented for rapid convergence. Mapping-based variable dividing is presented for precise related variables. Thirdly, based on minimum Manhattan distance, a triangle-approximating decision-making is proposed for assisting users' determination with/without preference information. Numerical experiments against several state-of-the-art feature selection algorithms demonstrate that the proposed DMEA-FS outperforms its competitors in terms of both classification accuracy and metrics of features' number.