Payment Fraud Detection using XGBoost
Model Description
This model predicts the probability of fraudulent transactions in payment networks using extreme gradient boosting (XGBoost).
Intended Use
- Primary use case: Real-time/batch flag generation for transaction risk scoring in payment gateways.
- Intended users: Risk analysis engines, data science teams evaluating tabular fraud pipelines.
Out-of-Scope Uses
- Do not use as an automated decision-maker for immediate account suspension without human review.
- Do not use on non-financial transactional datasets or credit scoring without retraining.
Training Data & Features
- Dataset trained on transaction attributes including velocity features, transaction amount, geographical risk scores, and device identifiers.
Evaluation Results
- Primary Metric (PR-AUC): Selected PR-AUC over accuracy due to extreme class imbalance (~0.1% fraud rate).
- PR-AUC: 0.89
- F1-Score: 0.84
Ethical Considerations & Limitations
- Bias & Fairness: Geographic and demographic features must be continuously audited for proxy discrimination.
- Drift: Performance requires monitoring against evolving fraud techniques and seasonal spending spikes.
- Downloads last month
- -