Learning from Positive and Unlabeled Examples: A Survey

This paper surveys the existing method of learning from positive and unlabeled examples. We divide the existing methods into three families, and review the main algorithms, respectively. The first Family of methods takes a two-step strategy, extracting some reliable negative examples, and then applying the supervised or semi-supervised learning method. The second family of methods estimates statistical queries over positive and unlabeled examples. The third family of methods reduces this problem to the problem of learning with high one-sided noise by treating the unlabeled set as noisy negative examples. Finally, we conclude and issue future works.