Fraud detection in transaction datasets

Abstract

Although the number of transaction fraud events grows slower than the number of transactions in total, it is still a problem for many institutions. Detecting fraudulent transactions is challenging for multiple reasons, including a general lack of labels, class imbalance, and hidden and evolving fraud patterns. Even more difficulties emerge while modeling public transaction datasets, namely feature anonymization, missing information, and data aggregation. This work suggests a pipeline of modeling fraudulent transactions, which accounts for most of those concerns based on other researchers’ experience. From the modeling approaches, one can distinguish those based on transaction features and those using graph anomaly detection methods. This research combines both methods and presents cross-validation results over two datasets. Performance scores did not indicate the superior predictive power of any presented approach. Nevertheless, the addition of graph features in the case of the second dataset significantly improved validation scores and therefore indicated the direction for further research.