Binary golden eagle optimizer combined with initialization of feature number subspace for feature selection
Xinkai Yang, Luhan Zhen, Zhanshan Li
Abstract
Feature Selection (FS) is a significant data preprocessing technique, whose purpose is to identify feature subset that can improve the prediction accuracy from the subsequent training model, while ensuring that the number of features is minimized. The Binary Golden Eagle Optimizer (BGEO) algorithm can efficiently search for the feature subset that satisfies the desired requirements on small and medium sized data. However, searching for subsets becomes much more difficult as the data dimension increases. To handle this problem, we present the Binary Golden Eagle Optimizer algorithm combined with Initialization of Feature Number Subspace (BGEO-IFNS) as the first application of BGEO for high-dimensional FS. Initialization of Feature Number Subspace (IFNS) consists of the following steps: First, the feature space is divided into several subspaces with different feature numbers. Next, new individuals are generated to correspond to the different subspaces. Finally, the quality of these new individuals is assessed using the original population. The initial population generated by this method improves the diversity while maintaining high quality, thus improving solving ability of the subsequent algorithm on high-dimensional data. To validate our approach, we conduct experiments on 14 small and medium dimensional datasets as well as 10 high-dimensional datasets. The experimental results show that BGEO-IFNS achieves a large improvement in effectiveness over BGEO and other state-of-the-art algorithms on the majority of datasets. In addition, BGEO-IFNS achieves optimal performance when compared with other metaheuristic algorithms utilizing IFNS. The results further validate the generalizability and adaptability of the new initialization method with BGEO.