An improved online multiple kernel classification algorithm based on double updating online learning

Abstract

Online multiple kernel classification(OMKC) algorithm is a promising algorithm in machine learning. Because of low error rate and relatively fast training time, it has been sucessfully applied to many real-world problems. However, in the phase of learning a single classifier for a given kernel, the OMKC adopts the perceptron algorithm, which significantly limits the performance of the algorithm. In this paper, we adopts the double updating online learning(DUOL) algorithm to learn the single classifier. Comparing to the perceptron algorithm, the DUOL algorithm not only assigns a weight to the misclassified example, but also updates the weight for one of the existing support vectors, which significantly improves the classification performance. Then we use the hedge algorithm to combines these classifiers. The experimental results show that the proposed algorithm is more effective than the OMKC algorithm, the state-of-the-art algorithms, and single kernel learning algorithm.