OBKA-FS: an oppositional-based binary kidney-inspired search algorithm for feature selection

Abstract

Feature selection is a key step when building an automatic classification system. Numerous evolutionary algorithms applied to remove irrelevant features in order to make the classifier perform more accurate. Kidney-inspired search algorithm (KA) is a very modern evolutionary algorithm. The original version of KA performed more effectively compared with other evolutionary algorithms. However, KA was proposed for continuous search spaces. For feature subset selection and many optimization problems such as classification, binary discrete space is required. Moreover, the movement operator of solutions is notably affected by its own best-known solution found up to now, denoted as Sbest. This may be inadequate if Sbest is located near a local optimum as it will direct the search process to a suboptimal solution. In this study, a three-fold improvement in the existing KA is proposed. First, a binary version of the kidney-inspired algorithm (BKA-FS) for feature subset selection is introduced to improve classification accuracy in multi-class classification problems. Second, the proposed BKA-FS is integrated into an oppositional-based initialization method in order to start with good initial solutions. Thus, this improved algorithm denoted as OBKA-FS. Third, a novel movement strategy based on the calculation of mutual information (MI), which gives OBKA-FS the ability to work in a discrete binary environment has been proposed. For evaluation, an experiment was conducted using ten UCI machine learning benchmark instances. Results show that OBKA-FS outperforms the existing state-of-the-art evolutionary algorithms for feature selection. In particular, OBKA-FS obtained better accuracy with same or fewer features and higher dependency with less redundancy. Thus, the results confirm the high performance of the improved kidney-inspired algorithm in solving optimization problems such as feature selection.