Evaluation of fraud detection data mining used in the auditing process of the Defense Finance And Accounting Service

Abstract

The Defense Finance and Accounting Service (DFAS) uses data mining to analyze millions of vendor transactions each year in an effort to combat fraud. The long timeline required to investigate potential fraud precludes DFAS from using fraud as a supervised modeling performance measure, so instead it uses the conditions needing improvement (CNI) found during site audits. To verify this method, a thorough literature review is conducted which demonstrates a clear relationship between fraud and CNIs. Then recent site audits are analyzed to prove that supervised modeling is detecting CNIs at a higher rate than random record selection. The next phase of the research evaluates recent models to determine if models are improving with each new audit. Finally, to enhance the upervised modeling process, four initiatives are proposed: a revised model scoring implementation, a knowledge base of audit results, alternative model streams for record selection and a recommended modeling process for the CNI knowledge base. The goal of the proposed enhancements is to improve an already successful program so that the data-mining efforts will further reduce taxpayer losses through fraud, error or misappropriation of funds.