A Dual Ensemble Learning Framework for Real-time Credit Card Transaction Risk Scoring and Anomaly Detection

Abstract

This paper presents a novel dual-entity learning method for real-time credit card fraud detection that combines advanced learning methods with dynamic risk methods. The framework employs a parallel processing architecture that combines XGBoost and deep ensemble models, enabling simultaneous transaction analysis through complementary detection streams. The system implements specialized feature engineering pipelines that generate 128 derived features through statistical transformations and domain-specific calculations. Our approach addresses the inherent class imbalance in credit card transaction data through adaptive sampling techniques and dynamic threshold adjustment mechanisms. The framework was analyzed using data on 284,807 transactions, including 492 fraud cases. The test results show the best performance with a detection accuracy of 99.96%, an accuracy of 99.95%, and a recovery of 98.91% while maintaining operational latencies below 25 milliseconds. The system achieves a 15% improvement in detection rate and a 35% reduction in false positives compared to traditional methods. The framework's definitions provide a specific explanation for fraud, ensuring compliance and operational transparency. Performance tests under various loads demonstrate the ability to perform well, achieving up to 5,000 changes per second while maintaining accuracy. The proposed system has created new benchmarks in detecting fraud while providing financial institutions with strategic solutions.