Pairwise constraint propagation-based transductive method for positive and unlabeled learning

Positive and Unlabeled (PU) learning has no labeled negative examples, so almost all of the semisupervised classification methods no longer work. In this paper, we propose a novel transductive method which can exploit the similarity and dissimilarity information of the examples simultaneously for PU learning. To obtain a novel regularizer, we combine spy technique and pairwise constraint propagation. The proposed regularizer makes the similar examples obtain similar labels while the dissimilar examples get different labels. Furthermore, our approach results in a convex quadratic programming problem which can be solved very easily. Experimental results on several benchmark datasets illustrate the feasibility and effectiveness of our approach. ©, 2015, Journal of Information and Computational Science. All right reserved.