Learning from Positive and Unlabeled Data with a Selection Bias