Multiple Kernel Learning via Ensemble Artifice in Reproducing Kernel Hilbert Space
Lu Kou, Jiajing Zhao, Jianming Zhang, Cheng Qin · 2020
In this paper, we propose a novel ensemble model of multiple kernels called Multiple Kernel Learning via Ensemble Artifice in Reproducing Kernel Hilbert Space (MKLEA) by minimizing unified ensemble loss in multiple Reproducing Kernel Hilbert Spaces (RKHSs). Different from the previous Multiple Kernel Learning (MKL) methods which attempt to seek a linear combination of basis kernels as a unified kernel; our MKLEA aims to find multiple solutions in corresponding RKHSs simultaneously. To achieve this goal, the multiple individual kernel losses are integrated into a unified ensemble loss. Therefore, each individual model can co-optimize to learn their optimal parameters in the ensemble framework. Experimental results on several UCI classification and computer vision datasets demonstrate that our proposed MKLEA model achieves best classification performances among comparative MKL methods, such as SimpleMKL, Matrix-Regularized MKL, SpicyMKL and GMKL.