Abstract
<title>Abstract</title> <p>Spatial crowdsourcing (SC) platforms require real-time predictive models to connect location-dependent tasks with mobile workers. The dynamic nature of SC, influenced by evolving factors like traffic and worker behaviour, often leads to data and concept drift that can impair task allocation. Current methods typically rely on binary drift detection and fixed strategies, which are inadequate for the rapidly changing SC environment. To tackle these issues, we propose DriftSense++, a dynamic framework for managing drift in SC. It incorporates advanced techniques, including entropy-based regional data-drift sensing, model-aware ADWIN with Adaptive Hoeffding Trees for concept drift detection, and an adaptive regional response policy. DriftSense++ features a closed-loop process for localised sensing, diagnosis, and recovery. We evaluated DriftSense++ on four real-world SC datasets (NYC Taxi, Yelp, Geolife, and Foursquare) under controlled synthetic drift scenarios spanning abrupt, gradual, mixed, and recurring patterns. Experimental results show F1-scores between 0.919 and 0.933 and false alarm rates around 10\%. DriftSense++ outperformed the best baseline (ADWIN) by 0.06-0.08 F1 points and reduced false alarms by 6–7 percentage points. Additionally, it decreases recovery time by 27–30% for recurring drift and improves F1-scores by 2.5-3.3% (relative) through dynamic partitioning. The adaptive response policy reduced unnecessary retraining by 4–5 updates per run without sacrificing recall. Overall, DriftSense++ significantly enhances drift management in SC systems by shifting from reactive detection to proactive, region-aware solutions.</p>