Analysis of Learning from Positive and Unlabeled Data