<![CDATA[Financial fraud detection in transaction streams faces distributional change, severe class imbalance, delayed outcome labels, and strict inference-time constraints. This study presents a drift-aware ensemble architecture in which ADWIN and Page–Hinkley statistics condition a hypernetwork that generates low-dimensional adaptation coefficients and detector-gating weights. Episodic meta-training is designed to expose the controller to simulated changes in fraud behavior, while a dual-timescale mechanism separates rapid configuration updates from slower controller refinement. The architecture further specifies versioned logging for decision traceability. An evaluation protocol is described for chronological transaction streams, including static and adaptive baselines, precision–recall performance, adaptation delay, serving latency, and component ablations. Numerical results presented in this manuscript require verification against the original experiment outputs and implementation before they can be treated as empirically established findings. The central contribution is a testable architecture for conditional ensemble adaptation; its comparative effectiveness and deployment readiness remain subject to reproducible validation.]]>
Copyrights © 2025
