Filter-based feature selection using two criterion functions and evolutionary fuzzification

Abstract

Real world problems often contain noise features which can decrease effectiveness of classification models. This article proposes a filter-based technique to select a minimal set of features for classification problems. The proposed method employs fuzzification of original features based on irregular-shaped membership functions created by genetic algorithm and particle swarm optimization, and a feature selection process using two criterion functions to evaluate feature subsets. The first function is applied to eliminate features with redundant effects, and the second function is applied to select a feature subset that maximizes inter-class distances and minimize intra-class distances. Standard machine learning data sets in various sizes and complexities are used in experiments. The results show that the proposed technique is effective and performs well in comparisons with other research.