In this paper, a novel ensemble kernel model based on game theory and repeated games, termed Repeated Game Kernel (RGK), is proposed to enhance the performance of Support Vector Machine (SVM). This model aims to maximize the advantages of both the polynomial kernel and the Radial Basis Function (RBF) kernel while achieving the highest compatibility with the given data. Unlike conventional kernel combination methods that may diminish the effect of one kernel, RGK strategically integrates both kernels in a staged manner to optimize neighborhood preservation and high-dimensional representation simultaneously. To enhance the model’s adaptability, a repeated game process is employed, wherein the RBF kernel’s radius is iteratively reduced and merged with the existing kernel. Experimental results demonstrate that RGK outperforms RBF, polynomial kernel, and previous ensemble kernels by achieving superior classification performance and greater adaptability to data distributions. Furthermore, increasing the number of game stages has been observed to strengthen the influence of the RBF kernel, leading to improved SVM accuracy in complex datasets.
