Learning from Double Positive and Unlabeled Data for Potential-Customer Identification