Real time credit card fraud detection on huge imbalanced data using meta-classifiers

Fraud detection in credit card transactions has several major challenges including the huge volume and high velocity of the transactions, data imbalance and frequent change in the fraud patterns. This paper presents a real-time tree based meta-classifier TBMC that can be used to identify fraudulent transactions in huge imbalanced data. The developed meta-classifier based model operates based on predictions in two levels. The first level of predictions is performed by Random Forest classifier, and the second level predictions are performed by an ensemble created with Decision Trees and Gradient Boosted Trees. The results obtained from first and the second level prediction models are integrated to form the final predictions. Experiments were conducted with UCSD-FICO data and the results were compared with state-of-the-art existing models, which showed high predictive levels of the developed TBMC model.