Feature subset selection for big data via parallel chaotic binary differential evolution and feature-level elitism

Abstract

Feature subset selection (FSS) employing a wrapper approach is fundamentally a combinatorial optimization problem maximizing the area under the receiver operating characteristic curve (AUC) of a classifier built on this subset under single objective environment. To balance both the AUC and the cardinality of the selected feature subset, we propose a novel multiplicative fitness function that combines AUC and a decreasing function of cardinality. Although the differential evolution algorithm is robust, it is prone to premature convergence, which can result in entrapment in local optima. To address this challenge, we propose chaotic binary differential evolution coupled with feature-level elitism (CE-BDE), where the chaotic maps are introduced at the initialization and the crossover operator. We also introduce feature-level elitism to improve the exploitation capability. Feature-level elitism involves preserving those features, which are chosen based on their frequency of occurrence in the population in the evolution process. Dealing with big data entails computational complexity, which motivates us to propose an effective parallel/ distributed strategy island model. The results demonstrate that the parallel CE-BDE outperformed the rest of the algorithms in terms of mean AUC and cardinality. The speedup and computational gain yielded by the proposed parallel approach further accentuate its superiority. Overall, the top-performing algorithm with the multiplicative fitness function turned out to be statistically significant compared to that with the additive fitness function across 5 out of 6 datasets.