Kolmogorov-Arnold networks guided whale optimization algorithm for feature selection in medical datasets

• DOI: 10.1186/s40537-025-01116-7
• Publication Date: 2025-03-21T19:29:48Z
• Authors: Boli Zheng; Yi Chen; Chaofan Wang; Ali Asghar Heidari; Lei Liu; Huiling Chen; Guoxi Liang
• Keywords: Feature (linguistics); Optimization algorithm

Abstract Feature selection (FS), dealing with pathological data with a high dimensionality and a small number of samples, has always been quite challenging. Among these, wrapper-based FS methods utilizing evolutionary algorithms have gained immense popularity due to their high accuracy. Additionally, utilizing surrogate models alongside wrapper-based FS methods can help save time by reducing the need for real model predictions. However, the use of surrogate models alongside evolutionary algorithms (EAs) presents two major challenges. Firstly, the training samples for the surrogate model all come from the iterative process of EAs. For surrogate models, the limited number of iterations in EAs results in a sparse amount of training samples. Additionally, FS emphasizes reducing feature quantity, which further contributes to the sparsity of the training data. In this paper, we propose a method for constructing training data based on the reliefF algorithm, which not only enables the acquisition of a large amount of training data but also helps in addressing the issue of feature sparsity. Furthermore, we propose the use of Kolmogorov-Arnold networks (KAN) as surrogate models to address the issue of sparse features. Finally, the whale optimization algorithm (WOA) was chosen as the evolutionary algorithm, as it has exhibited excellent performance in FS problems. Experimental results indicate that using KAN as a surrogate model to identify superior individuals generated by the WOA is feasible and has shown excellent performance in addressing medical FS problems.