Semi-supervised Online Multiple Kernel Learning Algorithm for Big Data

Abstract

In order to improve the performance of machine learning in big data, online multiple kernel learning algorithms are proposed in this paper. First, a supervised online multiple kernel learning algorithm for big data (SOMK_bd) is proposed to reduce the computational workload during kernel modification. In SOMK_bd, the traditional kernel learning algorithm is improved and kernel integration is only carried out in the constructed kernel subset. Next, an unsupervised online multiple kernel learning algorithm for big data (UOMK_bd) is proposed. In UOMK_bd, the traditional kernel learning algorithm is improved to adapt to the online environment and data replacement strategy is used to modify the kernel function in unsupervised manner. Then, a semi-supervised online multiple kernel learning algorithm for big data (SSOMK_bd) is proposed. Based on incremental learning, SSOMK_bd makes full use of the abundant information of large scale incomplete labeled data, and uses SOMK_bd and UOMK_bd to update the current reading data. Finally, experiments are conducted on UCI data set and the results show that the proposed algorithms are effective.