Context-aware Drift Detection for Quality-aware Data-Driven Smart City Digital Twins

Abstract

As urban environments become more complex, Smart Cities will increasingly depend on the integration of Internet of Things (IoT) devices and Digital Twin (DT) technologies to enable real-time monitoring, simulation, and predictive analytics, to ultimately improve quality of life. Data-driven approaches play a crucial role in optimizing city operations, but their effectiveness is hampered by data drift, shifts in data distributions over time that can degrade model performance. Frequent model retraining is a common solution, but it can be ineffective or lead to high computational costs. Alternatively, drift-driven approaches are often vulnerable to false detections caused by noisy IoT data, hardware failures, or cyber threats. To address that problem, we propose a context-aware drift detection algorithm that takes advantage of local correlations between data sources to detect actual real drifts while avoiding false positives. Furthermore, we present a decentralized layered architecture for embedding drift detection within the Smart City ecosystem and evaluate our approach by using a real-world dataset. Our approach demonstrates the viability of context-aware distributed drift detection, enhancing the reliability and efficiency of ML-driven Smart City applications.