Hyphatia: A Card-Not-Present Fraud Detection System Based on Self-Supervised Tabular Learning

In order to conduct credit card fraud, having only the payment card information of the victim it is possible to fake its identity and buy on e-commerce platforms. This type of fraud is known as Card-Not-Present and shows in the form of chargebacks, projecting billion-dollar losses worldwide in the coming years. The IEEE-CIS dataset has emerged as a strong option for creating and validating smart detection systems against this problem. In this work, we propose a solution, Hyphatia, where Self-Supervised Learning is implemented for tabular data based on SubTab. Our model outperforms XGBoost by 2.14% AUROC, detecting 67.44% of the fraud cases in the IEEE-CIS. This pioneering experimentation prioritizes those features that are not obfuscated. Furthermore, beyond providing just classification metrics, we also share time performance and feature importance calculations for explainability. To the best of our knowledge, this is one of the first works in the literature using Self-Supervised Learning for the problem of credit card fraud detection, specifically using the Self-Supervised Tabular Learning approach.