Hierarchical Uncertainty-Driven Adaptive Ensemble for Cost-Sensitive Stream Fraud Detection

Financial fraud detection faces a critical challenge: fraudulent patterns evolve continuously while traditional models remain static after training. This paper presents an adaptive ensemble framework combining incremental learning with hierarchical uncertainty quantification for streaming fraud detection under concept drift. The approach integrates Adaptive Random Forest, incremental SVM, online neural networks, and Bayesian networks with ADWIN-based drift detection and conformal prediction. Evaluated on IEEE-CIS dataset (590,540 transactions) with temporal validation, the framework achieves 92.3% accuracy and F1-score of 0.843 while adapting to distribution shifts within 1,053 samples. Calibrated uncertainty (ECE=0.074) is maintained through dynamic recalibration. The system processes transactions at 8.7ms median latency with 47.5% ROI compared to weekly retraining. Results demonstrate that temporal evaluation reveals 5-8% performance gaps masked by random splits, and incremental adaptation outperforms periodic retraining in cost-performance trade-offs.