Fuzzy filter cost-sensitive feature selection with differential evolution

Abstract

In recent years, a variety of feature selection methods based on evolutionary computation (EC) techniques have been developed for classification due to their robustness and search ability. However, the previous EC-based feature selection research mostly focuses on enhancing the prediction accuracy without taking the costs associated with the learning process into consideration. In other words, the impact of using EC-based feature selection to improve the cost-sensitive classification performance has not been well studied. Further, it is not possible to find a EC-based filter cost-sensitive feature selection method in the literature, to the best of our knowledge. We therefore design a cost-sensitive evaluation criterion using the principles of fuzzy mutual estimator and then adopt the criterion in differential evolution framework. According to a variety of experiments conducted on various benchmarks, the proposed filter method can effectively minimize both the classification error rate and the feature cost by removing irrelevant and distracting features from the dataset in a reasonable time.