A Clustering-based Multiple Kernel Learning Algorithm for Multi-Class Classification

Abstract

Multiple kernel learning algorithms typically optimize kernel alignment, structural risk minimization, and Bayesian functions. However, they have limitations, including inapplicability to multi-class classification, high time complexity, and no analytic solution. Analyzing clustering and classification similarities, we propose a novel clustering-based multiple kernel learning algorithm for multi-class classification (CBMKL). This algorithm transforms input space to high-dimension feature space using multiple kernel mapping functions. It estimates base kernel function weights and constructs the decision function using clustering objectives. This CBMKL algorithm has several advantages. 1) It handles multi-class problems directly. 2) This algorithm has an analytical solution, avoiding approximate solutions from sampling methods. 3) It also has polynomial time complexity. Experiments on two datasets illustrate these advantages.