A Novel Improved Binary Optimization Algorithm and Its Application in FS Problems

With the rapid advancement of artificial intelligence (AI) technology, the demand for vast amounts of data for training AI algorithms to attain intelligence has become indispensable. However, in the realm of big data technology, the high feature dimensions of the data frequently give rise to overfitting issues during training, thereby diminishing model accuracy. To enhance model prediction accuracy, feature selection (FS) methods have arisen with the goal of eliminating redundant features within datasets. In this paper, a highly efficient FS method with advanced FS performance, called EMEPO, is proposed. It combines three learning strategies on the basis of the Parrot Optimizer (PO) to better ensure FS performance. Firstly, a novel exploitation strategy is introduced, which integrates randomness, optimality, and Levy flight to enhance the algorithm’s local exploitation capabilities, reduce execution time in solving FS problems, and enhance classification accuracy. Secondly, a multi-population evolutionary strategy is introduced, which takes into account the diversity of individuals based on fitness values to optimize the balance between exploration and exploitation stages of the algorithm, ultimately improving the algorithm’s capability to explore the FS solution space globally. Finally, a unique exploration strategy is introduced, focusing on individual diversity learning to boost population diversity in solving FS problems. This approach improves the algorithm’s capacity to avoid local suboptimal feature subsets. The EMEPO-based FS method is tested on 23 FS datasets spanning low-, medium-, and high-dimensional data. The results show exceptional performance in classification accuracy, feature reduction, execution efficiency, convergence speed, and stability. This indicates the high promise of the EMEPO-based FS method as an effective and efficient approach for feature selection.