Hybrid Multi-Level Credit Card Fraud Detection System by Bagging Multiple Boosted Trees (BMBT)
M. S. Kavitha, M. Suriakala · 2017
The escalating usage levels of credit cards has resulted in the generation of enormous amounts of data, which in-turn has resulted in huge amounts of fraudulent activities in this domain. The major issues in fraud detection on credit card transaction data is that they are huge and they exhibit huge imbalance levels. This paper presents a multi-level ensemble model called Bagging Multiple Boosted Trees (BMBT) for credit card fraud detection by bagging multiple boosted trees. Bagging is performed by creating overlapping subsets of the training data. Each base learner of the bagged model is composed of boosted trees, enabling effective predictions by effectively handling data imbalance. Voting based prediction aggregation provides the final predictions. Experiments were conducted with UCSD-FICO data and comparisons with the existing models in literature indicates that the proposed BMBT model outperforms the existing models at the rate of 1%-5% in terms of AUC, BCR, BER and TNR levels.