Adaptive Confidence-Aware Transformer-Graph Fusion Network for Robust Credit Fraud Detection (Istanbul University - Journal of Electrical & Electronics Engineering)

Summary

• The ULB European cardholder dataset used in the study comprises 284 807 transactions.p.6
• Fraudulent transactions make up less than 0.2% of the ULB dataset, a highly imbalanced class distribution.p.7
• XGBoost achieved the best ROC-AUC among three machine learning models on the ULB dataset, with an AUC of 0.979.p.8
• XGBoost had the highest accuracy but an F1-score of only 0.28 at the selected threshold.p.9

Key insights

1. Confidence-aware late fusion combines transformer and GraphSAGE fraud probabilities using an adaptive weight.p.5
2. Fewer than 0.2% of transactions in the ULB dataset are fraudulent, so rare-event methods and metrics beyond accuracy are needed.p.7
3. Transaction amount alone is not a good indicator of fraud because the distributions of genuine and fraudulent amounts overlap substantially.p.8
4. XGBoost reached an AUC of 0.979 on the ULB dataset, the best of the three machine learning models compared.p.8
5. A high accuracy for XGBoost coexisted with a poor F1-score of 0.28, showing a weak balance between precision and recall.p.9