Smart Card Fraud Detection Using Machine Learning: A Comparative Study with Feature Engineering and Model Optimization

With the quick proliferation of the digital financial ecosystem, smart card fraud has seen a massive increase and caused billions of dollars of financial damage. In this paper, a comprehensive pipeline for smart card fraud detection over the real world and highly imbalanced dataset was developed. Three standard classifiers, including Logistic Regression (LR), Random Forest (RF), and XGBoost and the combination of RF and XGBoost were used. Rich feature engineering, including temporal features (hour, day, and time of the day), user behavior features (transaction count and amount deviation from average), merchant and category frequency encoding features, and geo distance features, was developed. The issue of class imbalance was handled using the scale_pos_weight parameter and thresholding in XGBoost. Hyperparameter tuning was carried out using RandomizedSearchCV and GridSearchCV to get the best performance. The tuned XGBoost model got 99.42% accuracy, 39.18% precision, 90.02% recall, and a 90.8% F-1 score. The above results indicate that a combination of rich feature engineering with state-of-the-art boosting methods can be a potent approach to combat real-world fraud.