Online semi-supervised support vector machine

Recently, support vector machine (SVM) has received much attention due to its good performance and wide applicability. As a supervised learning algorithm, the standard SVM uses sufficient labeled data to obtain the optimal decision hyperplane. However, in many practical applications, it is difficult and/or expensive to obtain labeled data. Besides, the standard SVM is a batch learning algorithm. It is inefficient to handle streaming data as the classifier must be retrained from scratch whenever a new data is arrived. In this paper, we consider the online classification of streaming data when only a small portion of data are labeled while a large portion of data are unlabeled. In order to obtain an adaptive solution with relatively low computational complexity, a new form of manifold regularization is proposed. Then, an adaptive and online semi-supervised least square SVM is developed, which well exploits the information of new incoming labeled or unlabeled data to boost learning performance. Simulations on synthetic and real data sets show that the proposed algorithm achieves good classification performance even if there only exist a few labeled data.