Comparative Evaluation of Multiple Kernel Learning Methods with Random Compact Gaussian Kernels

This work compares three kernel-based classification approaches: a Support Vector Machine with a random RBF kernel (RSVM); Multiple Random Compact Gaussian Kernel Learning (MRKL), with same weights (SW) or weights optimized by SimpleMKL (OP); and a Rational Function Kernel (RAFK), combining rational functions of dot products and Euclidean distances optimized with Optuna. The methods were evaluated on 15 benchmark datasets using 5-fold cross-validation and 20 independent runs. The MRKL variants achieved the highest mean accuracy on 13 of the 15 datasets (MRKL-OP on six and MRKL-SW on seven). RAFK produced promising results on some datasets, supporting the viability of rational-function kernels.