Li, Yu-Feng and Kwok, James and Zhou, Zhi-Hua
Semi-supervised support vector machines (S3VMs) are popular in using unlabeled data to help improve performance. Typical methods work by estimating the possible label assignments for unlabeled examples to maximize the margin, and generally suffer from some inefficiency issues. In this paper, in contrast to estimating the label assignment for unlabeled examples, we propose a new S3VM which works by estimating the label mean of unlabeled data and is therefore more efficient. We first show that S3VM with label means of unlabeled examples, referred to as meanS3VM, is nearly as same as supervised SVM with known labels for all unlabeled examples. Then, we present two algorithms which work by maximizing the margin between the label means of unlabeled data. Experiments show that, comparing with state-of-the-art methods, both of our proposed algorithms achieve highly competitive or even the best performance on semi-supervised benchmark tests, UCI data sets and newsgroup data. Moreover, our proposed algorithms are more efficient than existing S3VMs.
Discussion