A multi-objective differential evolution feature selection approach with a combined filter criterion

This paper proposes an improved filter evaluation criterion which uses the components of standard mutual information and fuzzy mutual information criteria by combining them in a simple and practical way. Then, a new filter approach is developed by integrating this criterion in multi-objective DE framework in order to enhance the performance in classification tasks. To verify the effectiveness of the developed filter approach, it is examined with single objective and multi-objective DE approaches based on both the standard mutual information and the fuzzy mutual information on a variety of benchmark datasets. The results indicate that the multi-objective DE filter approach based on the proposed filter criterion is able to achieve better classification accuracy and smaller feature subsets than other approaches based on existing criteria.