Adaptive Feature Selection and Ensemble Learning Framework for Multi-Domain Anomaly Detection in Real-Time Transactional Systems

Abstract

Anomaly detection in real-time transactional systems remains a critical challenge across financial, e-commerce, and digital advertising domains. Traditional approaches struggle with high-dimensional feature spaces, temporal dynamics, and cross-domain variability. This paper proposes an adaptive feature selection and ensemble learning framework that dynamically adjusts to evolving transaction patterns while maintaining computational efficiency. The framework integrates temporal behavioral analysis with multi-constraint optimization techniques to identify fraudulent activities across diverse operational contexts. Experimental results on multi-domain datasets demonstrate superior detection performance with 94.7% accuracy, 92.3% precision, and 91.8% recall, outperforming baseline methods by 12.4% in F1-score. The adaptive weighting mechanism reduces false positive rates by 34.6% compared to static ensemble approaches. The proposed framework achieves real-time processing latency under 45 milliseconds while maintaining detection quality across varying transaction volumes.