Feature Selection via GM-CPSO and Binary Conversion: Analyses on a Binary-Class Dataset

Abstract

Feature selection is oft-used to upgrade the system performance in classification-based applications. For this purpose, wrapper-based methods reserve an important place and are designed with efficient optimization methods so as to observe the highest performance. In this paper, a state-of-the-art optimization method named Gauss map-based chaotic particle swarm optimization (GM-CPSO) is handled. Binary conversion is considered to adapt the GM-CPSO to the feature selection. In classification part of the proposed method, k-nearest neighborhood (k-NN) is operated due to its fast and robust performance on classification-based implementations. In experiments, seven metrics (accuracy, sensitivity, specificity, g-mean, precision, f-measure, AUC) are utilized to objectively evaluate the performances, and 80%/20% training-test split is fulfilled to effectively assign the necessary features. Our wrapper-based method is tested on a balanced dataset that is based on Parkinson's disease (PD). As a result, our method presents promising scores by means of seven metrics, and especially, it improves the classification performance about 14.59% concerning the accuracy and AUC rates in comparison with the k-NN method.