A Comparative Study of Machine Learning Methods for Credit Card Fraud Detection

The problem of credit card fraud causes significant financial losses to businesses and consumers and prompts studies on effective ways of detecting fraud. Here we examine a publicly available dataset of credit card transactions (a total of 10,000) (151 fraudulent, 9,849 legitimate) and create supervised learning models to categorize credit card transactions. We train classifiers of logistic regression, random forest, and XGBoost, and the class imbalance will be addressed through the use of class-weighting. Performance is also assessed using accuracy, recall, F1-score, and ROC AUC (Receiver Operating Characteristics Area under the Curve) as sole accuracy will not suffice in the unbalanced data scenario. Evidence indicates that logistic regression achieves high recall (91 ) but low precision (23%), which is a high number of false alarm. Compared to the above, random forest is balanced (precision 100, recall 58, F1 =0.73), and the XGBoost is almost perfect in discrimination (precision 100, recall 97.8, F1 =0.99, ROC AUC =0.999). The results are consistent with available literature, which indicates that ensemble techniques (random forests and boosting) are more effective than simple models in detecting fraud. We speculate about the implications of this trade-off on a real-world deployment, as well as providing future work directions.