SMOTE-CHL-LightGBM: An Enhanced LightGBM for Credit Card Fraud Detection

Credit card fraud (CCF) is a worldwide risk for cardholders and financial institutions. The primary challenge in detecting CCF is dealing with extremely imbalanced data (EID). The EID problem presents two issues: an extremely low instance of fraud in training data and an extremely high risk of overlooking fraud instances. Augmenting the fraud data using oversampling can help with the first issue, but it is ineffective in addressing the latter. Cost-sensitive learning may be good for solving the second issue but is insufficient for the first one. This study proposes a novel method, SMOTE-CHL-LightGBM, to solve the EID problem comprehensively. The proposed method combines SMOTE, an oversampling technique, and cost-harmonization loss-based LightGBM, which was proposed in our earlier study. Experimental results indicate that the SMOTE-CHL-LightGBM outperforms baseline methods in critical metrics such as F2-score, AUC, recall, and cost savings. The proposed method shows great potential as a robust solution for financial applications facing EID problems.