Wrapper Feature Selection: Comparative Study of Binary Metaheuristic Algorithms

Feature selection is one of the major challenges in machine learning problems. It reduces the number of features and improves performance by removing the irrelevant and unnecessary ones. Finding the optimal subset of features in reasonable time is a difficult task, especially when the number of features is high. Feature selection can be considered as NP-hard problem in which the optimal solution can only be obtained through an exhaustive search. As an alternative, binary metaheuristics are effective tools for finding a sub-optimal solution for feature selection without testing all possible solutions. In this study, a comparative evaluation of eight binary metaheuristics, namely binary atom search optimisation (BASO), binary dragonfly algorithm (BDA), binary horse herd optimization algorithm (BHOA), binary grey wolf optimisation (BGWO), binary particle swarm optimization (BPSO), binary arithmetic optimization algorithm (BAOA), binary equilibrium optimizer algorithm (BEOA), and binary chimpanzee optimization algorithm (BChOA) for wrapper feature selection is conducted. The performance is evaluated on sixteen databases with various numbers of features and samples. The corresponding results demonstrate the effectiveness of feature selection in increasing accuracy, and BEOA appears as the best binary metaheuristic for wrapper feature selection.