Reducing false positives in bank anti-fraud systems based on rule induction in distributed tree-based models

Abstract

Fraud detection in bank payments transactions suffers from a high number of false positives. To deal with this problem, we introduce a rules generation framework for a fraud-detection system – an automatic rules generation using distributed tree-based ML (machine learning) algorithms such as Decision Tree, Random Forest and Gradient Boosting, where the components of expert rules are used as the features for the model. This approach is a combination of statistical and expert-based approaches. We apply it to the bank's card transaction data. Our dataset covers February 2021 and consists of more than 20 mil. records including information on clients, transactions, and merchants. The autogenerated rules were aimed at improving FPR (false positive rate) business-metric. The framework was tested in a real fraud-monitoring system of large bank throughout half of the year. The rules obtained using this framework proved to be satisfactory efficient while having tangible business effect.