Ensemble Machine Learning Frameworks for Real-Time Anomaly Detection in E-Commerce Transactions

Abstract

E-commerce platforms in the. U.S. incur $48 billion in annual fraud losses, projected to escalate 141% by 2029, disproportionately affecting resource-limited enterprises. This research proposes an ensemble machine learning framework for real-time anomaly detection, combining logistic regression for coefficient-based risk attribution with random forests for nonlinear feature robustness and isolation forests for unsupervised outlier identification. Key simulated fraud patterns include IP geolocation discrepancies, device-sharing indicators, and atypical purchase volumes. Trained on a synthetic corpus of 200,000 transactions (95% benign, 5% anomalous), the model attains 92% accuracy, 94% precision, and under 5% false positives-outperforming standalone approaches by a 15% increase in overall accuracy. Hyperparameter optimization using GridSearchCV improves predictive performance, while deployment on scalable cloud environments such as AWS EC2/S3 supports low-latency execution for real-time risk scoring and alerts. Scenario-based evaluations across varied transaction profiles highlight 7% improvements in fraud classification efficiency, with device-sharing emerging as a 75% risk amplifier. By promoting open-source dissemination, this framework supports broader adoption, particularly among resource-constrained enterprises. Based on projected transaction volumes and estimated fraud reduction rates, early adopters could avert $500 million financial losses. These projections are grounded in the ensemble's accuracy, real-time deployment capability, and identification of key risk amplification factors such as device-sharing. The framework also informs future enhancements like federated learning. Findings align with national priorities for resilient digital economies, emphasizing scalable AI for transactional integrity.