Multiobjective optimization algorithm with dynamic operator selection for feature selection in high-dimensional classification

Abstract

Feature selection (FS) is an important technique in data preprocessing that aims to reduce the number of features for training while maintaining a high accuracy for classification. In recent studies, FS has been extended to optimize multiple objectives simultaneously in classification. To better solve this problem, this paper proposes a new multiobjective optimization algorithm with dynamic operator selection for feature selection in high-dimensional classification, called FS-DOS. First, two complementary search operators with different characteristics are designed, where the first operator is a quick search (QS) operator aiming to accelerate the convergence speed, and the other operator is a modified binary differential evolution (BDE) operator that can prevent the algorithm from falling into a local optimum. In addition, a dynamic selection strategy based on the idea of resource allocation is also designed to dynamically select the most suitable operator for each solution according to its corresponding performance improvement on aggregated objective values. The simulation results on fifteen different real-world high-dimensional FS datasets show that FS-DOS can obtain a feature subset with higher quality than several state-of-the-art FS algorithms. Importantly, in terms of error rate, FS-DOS wins 55 out of 75 comparisons. In terms of dimensionality reduction, the number of features selected by FS-DOS is between one hundredth and one thousandth of the original dataset.