Machine Learning to Predict Government Fraud

Challenge

The General Services Administration (GSA) seeks to provide each executive agency that issues and uses purchase cards and convenience checks an effective system, techniques, and technologies to identify and prevent illegal, improper, or erroneous purchases. 

Guidehouse has been working with GSA, building dashboards that compare and explore spending profiles with historical data both within agencies and agencies of comparable size. However, to further enable the Inspector General  of each executive agency to complete required periodic risk assessments and audits of the agency purchase card or convenience check programs, as well as reduce the use of manual error-prone processes of individual transaction review, GSA had to extend its tool sets and leverage methods in machine learning, data mining, and associated technologies. 

Solution

Guidehouse leveraged its artificial intelligence and automation expertise by providing an automated and statistically unbiased solution for identifying patterns of transactions to be further reviewed by analysts. Our team implemented an isolation forest algorithm, a type of unsupervised machine learning algorithm, to help detect anomalies in the data that are flagged and prioritized for analyst review. Whenever our model detects anomalous transactions, meaning they diverge most from the common spending profiles, they are immediately queued for an analyst review as they are more likely to be cases of potential fraud and misuse.

Results

Our model was able to identify the specific days during which anomalous transactions are most prevalent, Thursdays and Fridays. Those two days represented 38% of total transactions yet accounted for 72% of anomalous transactions. We also identified that most suspicious transactions originated from purchase/travel cards rather than fleet cards, even though fleet cards make up the majority of transactions. This created a narrow and well-defined set of transaction behaviors to help single out and flag these transactions as anomalous, so that analysts might further inspect them to determine whether they are indeed fraudulent or not.