Experimental Study of the Instance Sampling Effect on Feature Subset Selection Using Permutational-Based Differential Evolution

Wrapper approaches for feature subset selection are computationally intensive because they require training and evaluation of a machine learning algorithm to assess the goodness of a subset of features. This proposal combines the permutational-based differential evolution for feature selection (DE-FSPM) algorithm as a wrapper approach with three instance sampling strategies: fixed, incremental, and evolving sampling fraction. These sampling schemes are applied to the search process to reduce the instance set used in an individual evaluation, resulting in overall computational time savings. In addition, the DE-FSPM algorithm is modified to have adaptive parameter control using success history-based parameter adaptation for differential evolution (SHADE). The experimental results show that using a reduced number of instances permits a reduction in computational cost with no significant differences in performance. The algorithm's use of adaptive parameter control did not improve its capabilities.