Helping a government agency to identify fraud
“Machine learning” yields big gain in detection accuracy
A Netherlands public agency responsible for providing social benefits was struggling to create an efficient system for detecting fraud in applications to receive social benefits. Only a very small fraction of the applications that were flagged by the system for possible fraud turned out to be actual fraud; in fact, the great majority was “false positives.” The need to investigate such a large number of false positives was overwhelming the agency’s capacity to review cases and root out real cases of deceit.
Deloitte Netherlands invited the agency to visit the Greenhouse for a demonstration of how advanced analytics could be used to detect fraud more reliably. The solution that Deloitte Netherlands’ analytics specialists had in mind was a “machine learning” program: one that would teach itself to flag benefits applications for possible fraud and use feedback from the fraud team’s analysis to gain accuracy over time.
Ahead of the Greenhouse session, Deloitte Netherlands received samples of the client’s actual benefits application records, some fraudulent and some legitimate. Working closely with the client’s software vendor, Deloitte Netherlands’ analytics team created a model to scan the records and devise a logical method of identifying applications that appeared fraudulent. That model would be put to the test in the Greenhouse.
Deloitte Netherlands began the Greenhouse session by explaining how machine learning works; in this case they showed how a computer learns to judge whether a head photographed from behind belongs to a man or a woman. The machine-learning model uses a “training” sample of photographs—photos identified as men or women—to create a formula for predicting gender based on a weighted combination of visual attributes such as hair length and neck size. Every successive training case enabled the model to gain prediction accuracy by changing the importance it assigns to each attribute.
Next, the team ran the fraud-detection model it had created for the client. The team fed a batch of application records into the model. Like the photographic-prediction model, the fraud-detection model became increasingly accurate as it churned through the new batch of records and refined its logic for making fraud predictions based on each record’s attributes. By the end of the session, the model was operating at approximately 75 percent prediction accuracy—a big improvement over the client’s existing accuracy rate.
The Greenhouse session provided the agency with a firsthand look at how predictive analytics really works. It was also the first time the agency had seen advanced analytics applied to their work, and it convinced them to proceed with implementing the predictive model that Deloitte Netherlands had showcased in the Greenhouse. The model was eventually flagging fraudulent applications with 95 percent accuracy. The model not only increased the efficiency of the fraud department, but also stopped a large number of fraudulent payments. Further analysis of the predicted and validated fraud cases also revealed valuable insight into fraud patterns.