AAI_2025_Capstone_Chronicles_Combined

The active learning simulation of Model 2 across different temporal update intervals reveals critical insights for production deployment strategies. Analysis of temporal update frequency demonstrates that longer retraining cycles provide superior stability, with 12-month updates maintaining consistent in-location accuracy between 60-90% throughout the simulation period, while 3-month updates exhibited high volatility with performance fluctuations ranging from 30-80%. This counter-intuitive finding suggests that frequent retraining on uncertain samples may destabilize model performance by overfitting to short-term noise rather than capturing meaningful temporal drift patterns. More significantly, despite a consistent 10-20% performance offset, out-location accuracy and F1-macro trends closely parallel in-location behavior across all update frequencies, demonstrating strong temporal transferability of active learning benefits across geographic domains. This correlated performance pattern indicates that temporal adaptation strategies learned from in-domain data effectively transfer to out-of-distribution geographic contexts, though with reduced absolute performance. These findings suggest that a single active learning strategy employing

296