Graph-Aided Online Multi-Kernel Learning
Authors: Pouya M. Ghari, Yanning Shen; 24(21):1−44, 2023.
Abstract
Multi-kernel learning (MKL) has gained popularity in function learning tasks. Unlike single kernel learning, which relies on a pre-selected kernel, MKL combines a dictionary of kernels to provide flexibility. However, including irrelevant kernels in the dictionary can decrease the accuracy of MKL and increase computational complexity. To address this challenge, a novel graph-aided framework is proposed to select a subset of kernels from the dictionary using a graph. Various graph construction and refinement schemes are developed based on incurred losses or kernel similarities to assist in the adaptive selection process. Additionally, to handle scenarios where data is collected sequentially or cannot be stored in batch due to its large scale, random feature approximation is adopted to enable online function learning. The proposed algorithms are proven to have sub-linear regret bounds. Experimental results on real datasets demonstrate the advantages of the novel graph-aided algorithms compared to state-of-the-art alternatives.
[abs]