EXPLAINABLE HYBRID FRAMEWORK FOR CREDIT CAR DFRAUD DETECTION USING TRANSFORMERS

Abstract

Credit card fraud causes huge losses for banks and customers every year. The problem is hard for three reasons: fraud is extremely rare compared to normal transactions, fraud patterns change over time, and regulators now require automated decisions to be explainable. Global card fraud reached $33.41 billion in 2024 and is projected to cross $403.88 billion over the next decade. This work proposes a single hybrid pipeline to address all three problems of the ULB Credit Card Fraud benchmark containing 284807 transactions and just 492 fraud cases (a 1:578 class ratio). The pipeline has six connected stages. A Taguchi L25 search picks a Quantile-then-Robust scaling pair, giving a five fold cross-validated AUC of 0.97940±0.00564. A Conditional Tabular GAN trained only on the 344 real fraud rows generates 132336 synthetic fraud samples, lifting the training set to a 40% fraud ratio. An autoencoder trained on legitimate transactions alone produces a reconstruction-error signal that separates fraud from non-fraud by a factor of 64 on the test set. A two-track design feeds augmented data to the tabular models and real-only chronological data to the sequence model. Three structurally different base learners, a Time-Aware Transformer (282,881 parameters), XGBoost on 64 leakage-free features, and an Explainable Boosting Machine, are combined through rank averaging. Platt scaling cuts test Expected Calibration Error from 0.498 to 0.0018, a 281-fold improvement, and three independent explainability methods (EBM shape functions, TreeSHAP, and LIME) unanimously rank autoencoder reconstruction error as the top predictor. On the held-out test set the system reaches accuracy of 0.9994, AUC-ROC of 0.9764, AUC-PR of 0.8244 (about 476 times the random baseline), F1 of 0.8088, precision of 0.8871, recall of 0.7432, and Matthews correlation coefficient of 0.8117 at the F1-maximising threshold; a high-recall operating point catches 86.5% of fraud while flagging only 1.6% of legitimate transactions. The work extends the author’s earlier ICCCNT 2025 conference work along five dimensions: generative oversampling, sequential modelling, glass-box interpretability, formal probability calibration, and three-method explainability cross-validation.