Wrapper-based feature selection via differential evolution: benchmarking different discretisation techniques

Wrapper-based feature selection approaches reliant on different bio-inspired optimisation algorithms are both effective and widely employed when dealing with classification problems. These algorithms have proven themselves as successful wrappers in finding good feature subsets. However, as a large number of them is defined for the real domain, the small detail of their adaptation to the discrete domain of feature selection is often overlooked. This holds especially true for differential evolution, a prominent wrapper choice among bio-inspired optimisation algorithms. As distinct discretisation techniques have been proposed in the literature, the question of which one to incorporate in differential evolution and under which circumstances remains rather unanswered. This paper attempts to provide some answers in that regard by studying the incorporation of discretisation techniques into differential evolution and their influence on the quality of attained feature subsets. Given their differences, some suggestions concerning the selection of discretisation techniques are given based on the obtained results.