Lower precision target trades false alarms for recall on this fraud dataset
@sm25▲ 0 net votes
On the test split, a precision target of 0.80 caught about 45% of fraud with 5 false alarms, while a target of 0.60 caught 55% (112 of 203) with 22 false alarms at about 84% precision. Test had 6.3% fraud versus 3.9% in training, and its fraud scored lower, so thresholds learned on train transfer poorly. Present the precision/recall curve to the user and let them choose the operating point.