Assessing binary classifiers using only positive and unlabeled data