Binary Differential Evolution-Based Feature Selection for Hand Gesture Classification

Abstract

In any classification problem involving a large feature set, the feature selection step is essential for not only reducing the computational complexity of the classifier by removing the redundant features, but also possibly improving the classification accuracy by utilizing the most relevant features. The utility of evolutionary algorithms such as differential evolution and genetic algorithm in feature selection has been shown in literature. However, unlike other evolutionary algorithms, including genetic algorithm, the use of differential evolution algorithm on a binary representation of the variables to be optimized is not straight forward, since the standard differential evolution needs to be modified. In this paper, a correlation-based feature subset selection strategy is formulated using a modified mutation operator given in literature for solving binary space optimization using differential evolution. The proposed feature subset selection strategy is applied to hand-activity classification using data collected from multiple surface electromyogram (sEMG) and motion sensors. The effect of tuning the crossover rate in the proposed algorithm on the feature selection process is studied. For the considered feature set, the proposed algorithm performs better as compared to genetic algorithm and other conventional approaches such as rank-based strategy ReliefF and the best first approach. The performance of the feature selection algorithms is compared in terms of the dimension of the selected feature subsets as well as the accuracies with which the hand activities are classified.