Online Semi-Supervised Learning with Multi-Kernel Ensemble

In many practical real-time applications, prediction functions should be learned online upon the examples arriving in sequence. It is usually infeasible to label all the examples in the stream. However, most of the state-of-art online learning methods that tackle the real-time prediction problem work are not able to exploit the unlabeled data. In this paper, an online semi-supervised learning method based on multi-kernel ensemble is proposed, which enables online learning even if the received example is unlabeled. This method exploits the compatibility of multiple learners from different RKHS over the unlabeled data, based on which a regularized instantaneous risk functional is derived. Online convex programming is employed to minimize the derived risk. The experimental results on UCI data sets and the application to network intrusion detection show that the proposed method can effectively exploit the unlabeled data to improve the performance of online learning.