Adaptive Learning Under Distributional Shift: A Controlled Evaluation of Static Versus Incremental Models for Rare-Event Detection in Simulated License Plate Recognition Streams

• Guruprasath Sankaran

  Author

Automatic License Plate Recognition (ALPR) systems deployed in urban environments process continuous, high-velocity data streams under inherently non-stationary conditions. However, most machine learning approaches assume stationary data and rely on offline evaluation, creating a gap between experimental validation and operational reality. This paper addresses this gap through a controlled comparison of static and adaptive learning strategies for rare-event detection in simulated ALPR streams. We develop a simulation testbed that independently controls operational scale, rare-event prevalence, and temporal data drift. Six supervised models—three static (Logistic Regression, Random Forest, HistGradientBoosting) and three adaptive incremental learners (Hoeffding Tree, Adaptive Random Forest, Leveraging Bagging)—are evaluated using a prequential test-then-train protocol across class prevalence levels of 1–5% with controlled drift. Results show that adaptive models achieve F1-scores (0.927–0.930) comparable to static models (0.921–0.928), with overlapping 95% confidence intervals, indicating no significant accuracy advantage under stable conditions. Computational analysis reveals important trade-offs: Hoeffding Tree processes 11,920 instances per second—747× faster than Random Forest—with only a 0.6% F1 reduction, while Adaptive Random Forest offers a balanced compromise between accuracy and throughput. Under simulated distribution shift, adaptive models generate substantially fewer false positives than static models, demonstrating greater robustness. These findings provide practical guidance for selecting learning architectures for privacy-preserving, real-time ALPR systems operating under evolving data distributions.

• PDF