A Two-Stage Hybrid Fraud Detection Framework with Learnable Fusion and Probability Calibration for Credit Card Transactions

Credit card fraud detection is still a major problem in financial systems because of the extreme class imbalance, changing nature of fraud and the high cost of misclassification. Here we propose a novel two- stage hybrid detection framework that combines a supervised XGBoost classifier and an unsupervised Autoencoder-based anomaly detector. The outputs of both stages are fused by a learnable Logistic Regression meta-model that is trained on out-of-fold (OOF) predictions to prevent data leakage. We use Platt scaling to calibrate the fusion output so that the risk quantification is reliable. The calibrated fraud probability is then translated into a three tier risk bucketing scheme — LOW, MEDIUM, and HIGH — providing for auditable, actionable decisions. Our method outperforms the single-model baselines with a test ROC-AUC of 0.9776, PR-AUC of 0.8186, fraud precision of 0.99, recall of 0.72 and an overall accuracy of 99.97% on the publicly available European Credit Card dataset. We further build an operational dashboard with Streamlit for real- time inference and batch analysis, visualization of behavioural patterns and model health monitoring.

Keywords: Credit card fraud detection, XGBoost, autoencoder, anomaly detection, ensemble learning, SMOTE, probability calibration, risk stratification, out-of-fold stacking.

Chitra L, Gopika K L, Jasnavi N, Kavya S, Vaijayanthi M (2026). A Two-Stage Hybrid Fraud Detection Framework with Learnable Fusion and Probability Calibration for Credit Card Transactions. International Journal of Innovative Research in Multidisciplinary Education & Technology (IJIRMET), 11(3).