GAAMmf: genetic algorithm with aggressive mutation and decreasing feature set for feature selection

Abstract

This paper introduces a modified version of a genetic algorithm with aggressive mutation (GAAM), one of the genetic algorithms (GAs) used for feature selection. The modification proposed in this study expands the original GAAM’s capabilities by allowing not only feature selection but also feature reduction. To obtain this effect, we applied the concept of ranks used in the non-dominated sorting genetic algorithm (NSGA) and the concept of penalty term used in the Holland genetic algorithm. With those two concepts, we managed to balance the importance of two competing criteria in the GAAM fitness function: classification accuracy and the feature subset’s size. To assess the algorithm’s effectiveness, we evaluated it on eleven datasets with different characteristics and compared the results with eight reference methods: GAAM, Melting GAAM, Holland GA with a penalty term, NSGA-II, Correlation-based Feature Selection, Lasso, Sequential Forward Selection, and IniPG (an algorithm for particle swarm optimisation). The main conclusion drawn from this study is that the genetic algorithm with aggressive mutation and decreasing feature set (GAAMmf) introduced in this paper returned feature sets with a significantly smaller number of features than almost all reference methods. Furthermore, GAAMmf outperformed most of the methods in terms of classification accuracy (except the original GAAM). In contrast to Holland GA and NSGA-II, GAAMmf was able to perform the feature reduction task for all datasets, regardless of the initial number of features.