Differential evolution (DE) for multi-objective feature selection in classification

Feature selection has two main conflicting objectives, which are to minimise the number of features and maximise the classification accuracy. Evolutionary computation techniques are particularly suitable for solving mult-objective tasks. Based on differential evolution (DE), this paper develops a multi-objective feature selection algorithm (DEMOFS). DEMOFS is examined and compared with two traditional feature selection algorithms and a DE based single objective feature selection algorithm. DEFS aims to minimise the classification error rate of the selected features. Experiments on nine benchmark datasets show that DEMOFS can successfully obtain a set of non-dominated feature subsets, which include a smaller number of features and maintain or improve the classification performance over using all features. In almost all cases, DEMOFS outperforms DEFS and the two traditional feature selection methods in terms of both the number of features and the classification performance.