A binary water wave optimization for feature selection

Abstract

A search method that finds a minimal subset of features (over a feature space) that yields maximum classification accuracy is proposed. This method employs rough set theory (RST) along with a newly introduced binary version of the water wave optimization approach (WWO) which is denoted by BWWO. WWO simulates the phenomena of water waves, such as propagation, refraction, and breaking and is one of the newest nature inspired methods for global optimization problems. In our approach, BWWO utilizes the phenomena of water waves propagation, refraction, and breaking in a binary version. Two main experiments based on the rough set approach and wrapper method as a part of the objective function are carried out to verify the performance of the proposed algorithm. In the first experiment, the effectiveness of the proposed approach based on RST is demonstrated on 16 different datasets. The proposed approach is compared with various typical attribute reduction methods and popular optimizers in the literature, such as ant colony, nonlinear great deluge algorithm, scatter search and others. For the second experiment, a feature subset that maximizes the classification accuracy (using cross-validated kNN classifier) with minimizing the number of selected features is obtained over 17 different datasets. In wrapper experiment BWWO is compared with the binary gray wolf optimization, binary particle swarm optimizer, binary cat swarm optimization, binary dragonfly algorithm and the binary bat algorithm. The computational results demonstrate the efficiency and effectiveness of the proposed approach in finding a minimal features subset that maximize the classification accuracy. Furthermore, Friedman test and Wilcoxon's rank-sum test are carried out at 5% significance level in this study.