A Supervised Auto-Tuning Approach for a Banking Fraud Detection System

摘要

In this paper, we propose an extension to Banksealer, one of the most recent and effective banking fraud detection systems. In particular, until now Banksealer was unable to exploit analyst feedback to self-tune and improve its performance. It also depended on a complex set of parameters that had to be tuned by hand before operations. To overcome both these limitations, we propose a supervised evolutionary wrapper approach, that considers analyst's feedbacks on fraudulent transactions to automatically tune feature weighting and improve Banksealer's detection performance. We do so by means of a multi-objective genetic algorithm. We deployed our solution in a real-world setting of a large national banking group and conducted an in-depth experimental evaluation. We show that the proposed system was able to detect sophisticated frauds, improving Banksealer's performance of up to 35% in some cases.