Estimating the F1 Score for Learning from Positive and Unlabeled Examples