Stochastic fractal equilibrium optimizer with X-shaped dynamic transfer function for solving large-scale feature selection problems

Abstract

Large-scale feature selection (FS) is an important task in the field of data extraction and machine learning. Its core goal is to identify and screen out the feature subset that is most critical to the prediction target from the extensive initial attribute collection, in a bid to improve the efficiency and generalization of the model and reduce the computational complexity. A stochastic fractal equilibrium optimizer based on X-shaped dynamic transfer function is targeted for large-scale feature selection. Among them, the X-shaped transfer function is designed based on the scaling and flip changes of the basic transfer functions, and then the parameters related to the number of iterations in the equilibrium optimizer (EO) are further used as time-varying factors to dynamically change the X-shaped transfer function. The flower-shaped transfer function and dynamic flower-shaped transfer function are extended. At the same time, EO is optimized and improved by using the diffusion and updating process in the stochastic fractal search. To assess the efficacy and dominance of the designed FS method, 20 large-scale datasets were picked from UCI datasets for experiments, and SFEO-TF with the best performance was selected and set against 8 other intelligent optimization approaches. The experimental results show that the designed dynamic time-varying X-shaped transfer function is effective as well as the EO based on stochastic fractal search. Among them, in the two groups of controlled experiments, SFEO-TF has the best performance in performance metric and categorization precision, and even the performance of SFEO-TF set against 8 other intelligent optimization approaches in the minimum fitness value of all 20 large-scale datasets has reached the minimum value.