Anomaly Detection

A fraud detector that catches 80% of the fraud and wrongly flags 1% of legitimate transactions would sound respectable on a vendor's slide. Run it over the 284,807 card transactions that European cardholders made in two days of September 2013, which the Machine Learning Group at the Université Libre de Bruxelles later published as a benchmark. 492 of them were fraudulent, 0.172%. The detector raises about 394 true alerts and 2,843 false ones, so an analyst who opens an alert finds fraud about one time in eight.