Multiple Random Subset-Kernel Learning

摘要

In this paper, the multiple random subset-kernel learning (MRSKL) algorithm is proposed. In MRSKL, a subset of training samples is randomly selected for each kernel with randomly set parameters, and the kernels with optimal weights are combined for classification. A linear support vector machine (SVM) is adopted to determine the optimal kernel weights; therefore, MRSKL is based on a hierarchical SVM. MRSKL outperforms a single SVM even when using a small number of samples (200 to 400 out of 20,000 training samples), while the SVM requires more than 4,000 support vectors.