Transfer Learning Strategies for Credit Card Fraud Detection

Credit card fraud jeopardizes the trust of customers in e-commerce transactions. This led in recent years to major advances in the design of automatic Fraud Detection Systems (FDS) able to detect fraudulent transactions with short reaction time and high precision. Nevertheless, the heterogeneous nature of the fraud behavior makes it difficult to tailor existing systems to different contexts (e.g. new payment systems, different countries and/or population segments). Given the high cost (research, prototype development, and implementation in production) of designing data-driven FDSs, it is crucial for transactional companies to define procedures able to adapt existing pipelines to new challenges. From an AI/machine learning perspective, this is known as the problem of transfer learning . This paper discusses the design and implementation of transfer learning approaches for e-commerce credit card fraud detection and their assessment in a real setting. The case study, based on a six-month dataset (more than 200 million e-commerce transactions) provided by the industrial partner, relates to the transfer of detection models developed for a European country to another country. In particular, we present and discuss 15 transfer learning techniques (ranging from naive baselines to state-of-the-art and new approaches), making a critical and quantitative comparison in terms of precision for different transfer scenarios. Our contributions are twofold: (i) we show that the accuracy of many transfer methods is strongly dependent on the number of labeled samples in the target domain and (ii) we propose an ensemble solution to this problem based on self-supervised and semi-supervised domain adaptation classifiers. The thorough experimental assessment shows that this solution is both highly accurate and hardly sensitive to the number of labeled samples.