An Improved Unsupervised Modeling Methodology For Detecting Fraud In Vendor Payment Transactions
Abstract
In this thesis, we propose a standardized procedure for detecting fraud in Defense Finance and Accounting Service (DFAS) vendor payment transactions through Unsupervised Modeling (cluster analysis) . Clementine Data Mining software is used to construct unsupervised models of vendor payment data using the K-Means, Two Step, and Kohonen algorithms. Cluster validation techniques are applied to select the most useful model of each type, which are then combined to select candidate records for physical examination by a DFAS auditor. Our unsupervised modeling technique utilizes all the available valid transaction data, much of which is not admitted under the current supervised modeling procedure. Our procedure standardizes and provides rigor to the existing unsupervised modeling methodology at DFAS. Additionally, we demonstrate a new clustering approach called Tree Clustering, which uses Classification and Regression Trees to cluster data with automatic variable selection and scaling. A standardized procedure for Unsupervised Modeling, detailed explanation of all Clementine procedures, and implementation of the Tree Clustering algorithm are included as appendices.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2003
- Accession Number
- ADA417599
Entities
People
- Gregory W. Rouillard
Organizations
- Naval Postgraduate School