ABSTRACT

In the faceless world of the Internet, online fraud is one of the greatest reasons of loss for web merchants. Advanced solutions are needed to protect e-businesses from the constant problems of fraud. Many popular fraud detection algorithms require supervised training, which needs human intervention to prepare training cases. Since it is quite often for an online transaction database to have Terabyte-level storage, human investigation to identify fraudulent transactions is very costly. This paper describes the automatic design of user profiling method for the purpose of fraud detection. We use a FP (Frequent Pattern) Tree rule-learning algorithm to adaptively profile legitimate customer behavior in a transaction database. Then the incoming transactions are compared against the user profile to uncover the anomalies. The anomaly outputs are used as input to an accumulation system for combining evidence to generate high-confidence fraud alert value. Favorable experimental results are presented.

Keywords: Fraud detection, FP tree, anomalies, adaptive mining, association mining