Research on the ensemble feature selection algorithm based on multimodal optimisation techniques

Abstract

Feature selection is essentially a high-dimensional combinatorial optimisation problem. To find representative feature subsets, the selection method needs powerful exploration ability. In addition, if alternative feature subsets could be provided, the final prediction accuracy can be improved by ensembling these subsets. Multimodal optimisation (MO) methods with high exploration power can find multiple suitable solutions in one single run. Therefore, this paper presents an ensemble feature selection algorithm based on multimodal optimisation techniques. Differential evolution based on fitness Euclidean-distance ratio (FERDE) algorithm is utilised to search for multiple diverse feature subsets in the huge feature space. A set of diverse-based classifiers are built based on these subsets and ensemble to improve the final classification performance. Compared with several existing classical algorithms and ensemble feature selection methods, the proposed method can achieve higher predictive accuracy.