A novel quantum grasshopper optimization algorithm for feature selection

Abstract

Feature selection is an indispensable work to make the data mining more effective. It reduces the computational complexity and effectively improves the performance of learning models. The exhaustive algorithm and the greedy algorithm cannot adapt to the current increasing number of features when finding the potential optimal feature subset. Therefore, the feasible way for feature selection called swarm intelligence algorithms becomes popular. The grasshopper optimization algorithm is a novel swarm intelligence algorithm which has good performance. In this study, we improve the grasshopper optimization algorithm by applying quantum method. A dynamic population quantum binary grasshopper optimization algorithm based on mutual information and rough set theory (DQBGOA_MR) for feature selection is proposed. Through the quantization of grasshopper individuals, the search scope of feature space is improved, and a good balance is achieved between exploration and exploitation. The premature and catastrophe strategies are used to avoid converge prematurely and fall into a local optimum. Moreover, the rough set and mutual information based evaluation criterion is defined which considers both effectiveness of selected features and the relation among unselected features, selected features, and classes. The proposed method is evaluated by extensive experiments in twenty UCI datasets. Experimental results show that DQBGOA_MR has better feature reduction rate, higher classification accuracy and more stable results compared to other swarm intelligence algorithms.