A multi-objective feature selection approach based on chemical reaction optimization

Abstract

Feature selection (FS) is an NP-hard combinatorial optimization problem, which aims to select the most relevant features from a large number of candidates. Recently, the FS in classification has been handled as a bi-objective optimization problem, where both the classification error and the number of selected features are minimized simultaneously. Despite of the effectiveness of existing multi-objective optimization (MOO) algorithms on FS, the convergence toward the Pareto front and the distribution over the Pareto front still deserve to be improved. To this end, this paper proposes a new multi-objective chemical reaction optimization algorithm for FS, termed MOFSCRO. In the proposed MOFSCRO, the preference information related to the problem is firstly extracted from the population. Then, three evolutionary operators based on the preference information are suggested to guide the population evolution, with the purpose of enhancing the population convergence and diversity. Experimental results on 12 benchmark data sets demonstrate the superiority of the proposed MOFSCRO over seven state-of-the-art FS algorithms in terms of both the classification error and the number of selected features. In comparison to five MOO algorithms, MOFSCRO also shows better performance according to the metric of inverted generational distance.