Unsupervised Machine Learning for Card Payment Fraud Detection

Credit card fraud is one of the most common cybercrimes experienced by consumers today. Machine learning approaches are increasingly used to improve the accuracy of fraud detection systems. However, most of the approaches proposed so far have been based on supervised models, i.e., models trained with labelled historical fraudulent transactions, thus limiting the ability of the approach to recognise unknown fraud patterns. In this paper, we propose an unsupervised fraud detection system for card payments transactions. The unsupervised approach learns the characteristics of normal transactions and then identify anomalies as potential frauds. We introduce the challenges on modelling card payment transactions and discuss how to select the best features. Our approach can reduce the equal error rate (EER) significantly over previous approaches (from \(11.2\%\) to \(8.55\% ERR\)), for a real-world transaction dataset.