Semi-Supervised Classification Based on Classification from Positive and Unlabeled Data

Most of the semi-supervised learning methods developed so far use unlabeled data for regularization purposes under particular distributional assumptions such as the manifold assumption. On the other hand, recently developed methods of learning from positive and unlabeled data (PU learning) use unlabeled data for loss evaluation, i.e., label information is directly extracted from unlabeled data. In this paper, we extend PU learning to also incorporate negative data and propose a novel semi-supervised learning approach. We establish a generalization error bound for our novel method and show that the bound decreases with respect to the number of unlabeled data without the distributional assumptions that are required in existing semi-supervised learning methods. Through experiments, we demonstrate the usefulness of the proposed method.