Robust Anomaly Detection: SNAP Clustering Patterns

J. Woody
Mississippi State University, Mississippi, United States

Poster stand number: T112

Keywords: Anomaly Detection, Program Integrity, Fraud Detection, Fiancial Data

Food insecurity is a national security imperative. With worldwide declines in food production and distribution resulting from years of stress induced by the Covid-19 pandemic in addition to geopolitical disruptions, nutritional assistance should be poised to adapt to new stresses. Our team consisting of statisticians and data scientist from Google, the University of California at Santa Cruz, and Mississippi State University propose novel statistical method to identify anomalous activity in the Supplemental Nutritional Assistance Program (SNAP). We provide novel statistical methods which identify problematic activity and rank order recommendations for investigations. These methods can harden SNAP program to organized attempts to defraud the program. Our methods are the result of years of study of Electronic Benefit Transfer transactions in the state of Mississippi. While originally designed for the federal SNAP program, our methods and expertise will translate to anomaly detection in PPP loans, healthcare fraud, and inadequate remote worker performance. Using an elegant extension to Campbell's Theorem, we connect order statistics to Poisson Processes. Our methods can detect problematic patterns in domains as varied as transaction amounts, events in time, or (with adjustments) even spatial data.