Improved cost-sensitive learning for financial fraud detection and prevention
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Credit card fraud increases as e-commerce becomes more prevalent. Current fraud prevention and detection techniques, however, are far from accurate and effective. The low adoption rate of secondary verification techniques and high false positive declines lead to significant financial losses to merchants and card issuers and, at the same time, inconveniences and dissatisfactions to consumers. In this thesis, we provide effective solutions to improve both credit card fraud prevention and detection. For the fraud prevention, we introduce a transaction-value-based consumer incentive strategy for a merchant to adopt more accurate fraud prevention techniques, such as secondary verification. We identify the conditions under which the strategy is attractive to a merchant and show that it may lead to a `win-win-win' for consumers, the merchant, and the card issuer. For the fraud detection, we propose a framework that systematically calibrates the fraud probability estimate accuracy for each credit card transaction and design three cost-sensitive decision-making algorithms to minimize various fraud detection costs. We also propose a universal benchmark algorithm for a threshold-based model, which calculates the lower bound of the overall model cost and the corresponding threshold setting for a given testing dataset. The proposed algorithms are shown to be more effective than existing techniques via numerical tests using a real credit card transaction dataset from a European bank.
Description
Keywords
Consumer incentive, Cost-sensitive learning, Credit card fraud detection, Optimization,