An adaptive profile based fraud detection framework for handling concept drift

As e-commerce continues to grow, so does the opportunity for perpetrating online fraud. As a result many researches have been conducted to make online transactions possible in a risk free environment by proposing different fraud detection methods. Concept drift is an inherent feature in many data streams such as electronic financial transactions. Hence, many fraud detection techniques have tried to detect and preferably manage concept drift. In this paper, a new concept drift management framework has been proposed. In this framework a temporary profile has been introduced in order to retain new concepts in the incoming data stream independently from historical profile. When the historical profile reaches a different decision from the temporary profile this is an indication that most probably a concept drift has occurred. In this case, a window based method is applied as a strategy for managing concept drift. The ability to adapt normal profiles systematically makes this concept drift management framework applicable to any profile based fraud detection method. Simulation results indicate that the proposed scheme is able to reduce the false positives (FPs) of a typical fraud detection method to 4.3% on average in the presence of a wide variety of concept drifts in the incoming transactions. This is an average of 85.7% reduction in FPs for this fraud detection technique.