Positive and unlabeled (PU) learning is a fundamental task in many applications, which trains a binary classifier from only PU data. Existing PU learning methods typically assume that training and ...