Feature Subset Selection Using a Self-adaptive Strategy Based Differential Evolution Method

Feature selection is a key step in classification task to prune out redundant or irrelevant information and improve the pattern recognition performance, but it is a challenging and complex combinatorial problem, especially in high dimensional feature selection. This paper proposes a self-adaptive strategy based differential evolution feature selection, abbreviated as SADEFS, in which the self-adaptive elimination and reproduction strategies are used to introduce superior features by considering their contributions in classification under historical records and to replace the poor performance features. The processes of the elimination and reproduction are self-adapted by leaning from their experiences to reduce search space and improve classification accuracy rate. Twelve high dimensional cancer micro-array benchmark datasets are introduced to verify the efficiency of SADEFS algorithm. The experiments indicate that SADEFS can achieve higher classification performance in comparison to the original DEFS algorithm.