XGBoost-Based Online Payment Fraud Detection System: From Data Imbalance to Production Deployment

This article introduces an end-to-end online payment fraud detection project. It corely uses the XGBoost algorithm, adopts SMOTE oversampling and scale_pos_weight parameter tuning to address the problem of highly imbalanced financial data, optimizes recall rate, and implements production deployment via Streamlit. It forms a complete closed loop from data processing to model application, providing a practical solution for financial risk control.

With the popularity of digital payments, fraud detection is a core risk control capability for financial institutions, but it faces the problem of class imbalance (fraudulent transactions often account for less than 1%). If only overall accuracy is pursued, the model tends to predict all transactions as normal, losing practical value. Therefore, the project focuses on recall rate, prioritizing ensuring no fraudulent transactions are missed, even if a higher false positive rate is accepted.

To address data imbalance, the project adopts a dual strategy:

1. SMOTE Oversampling: Generate synthetic fraud samples to balance training data, avoiding overfitting caused by simple duplication;
2. scale_pos_weight Parameter: Adjust the weight of positive and negative samples in XGBoost without modifying the original data, which is flexible and efficient. In terms of feature engineering, it is designed based on business logic: account balance changes, transaction amount deviation from history, transaction frequency time distribution, payee behavior patterns, etc. Key insight: Fraud detection relies on behavior patterns rather than simple transaction amounts (small and frequent transactions may be more risky).

The project compares three algorithms: logistic regression (baseline, strong interpretability), random forest (stable integration, conducive to feature analysis), and XGBoost (final choice, gradient boosting framework that supports custom loss and weight adjustment). Optimization strategies include:

• Threshold Tuning: Analyze the precision-recall curve to select the optimal threshold, tending to lower the threshold to improve recall rate;
• Evaluation Metrics: Focus on recall rate (proportion of fraudulent transactions identified), precision (proportion of true fraud among predicted fraud), and F1 score (comprehensive indicator).

The final XGBoost model's performance on the test set:

Metric	Value	Interpretation
Recall Rate	~0.68	Identifies 68% of fraudulent transactions
Precision	~0.24	24% of predicted fraud are true fraud
F1 Score	~0.35	Comprehensive performance
Business value: Intercepts most fraudulent transactions; although the false positive rate is high, the cost of manual review is lower than the loss from fraud, so it can be used as the first layer of screening for manual review.

The project builds a web application via Streamlit, with features including real-time prediction of single transactions, fraud probability visualization, and batch data upload prediction. Streamlit advantages: Pure Python development, no front-end experience required, high efficiency. Deployment solutions: Already launched on Streamlit Community Cloud (suitable for prototype demonstration and lightweight production); for enterprise-level, consider Docker containerization, Kubernetes cluster, or integration with existing risk control systems.

Experience Summary: 1. Business understanding takes priority (defining optimization goals and balancing automation and manual review are more important than parameter tuning); 2. Imbalanced data requires flexible combination of strategies such as SMOTE, weight adjustment, and threshold tuning; 3. Need to complete production links such as model persistence, feature pipeline encapsulation, and web deployment. Improvement Directions: Improve recall rate (ensemble learning/deep learning), strengthen real-time feature engineering, increase model interpretability (SHAP/LIME), and support online learning to adapt to the evolution of fraud patterns.