Real-Time Transaction Fraud Detection Using Ensemble Learning and Graph Neural Networks

Financial fraud detection requires scalable and interpretable solutions capable of operating in real time. This work presents a production-oriented fraud detection framework integrating Isolation Forest, XGBoost, and Graph Neural Networks within a weighted ensemble architecture. The proposed system detects behavioral anomalies and relational fraud patterns while maintaining interpretability using SHAP-based explanations. A concept drift monitoring mechanism ensures long-term robustness against evolving fraud strategies. Experimental evaluation demonstrates a ROC-AUC of 0.94 with sub-250ms inference latency, highlighting the framework’s suitability for real-world financial deployment