Human-Centered Urban Intelligence through Context-Aware Recommendation Using IoT and Digital Twins

Abstract

Rapid urbanization has increased the need for intelligent urban systems that can support context-sensitive decisions for individuals operating under changing environmental and infrastructural conditions. This thesis addresses the lack of an integrated operational framework that connects Internet of Things sensing, artificial intelligence, and Digital Twin representations for personalized services in smart urban environments. The research is motivated by the fragmentation of existing smart-city systems, in which sensing, analytics, recommendation, and simulation are often developed as separate components and are commonly evaluated mainly through accuracy-oriented metrics.

To address this gap, the thesis proposes a unified framework that combines IoT-based context acquisition, structured context modeling, context-aware predictive and recommender models, and interoperable Digital Twins of users, points of interest, and urban environments. The framework transforms heterogeneous urban data into multi-dimensional contextual representations spanning environmental, spatial, temporal, and activity-related factors. It employs a hybrid wide-and-deep learning architecture to support prediction and recommendation under dynamic conditions. Within the proposed pipeline, Digital Twins extend recommendations beyond immediate ranking by enabling simulation-based reasoning over alternative actions and their anticipated consequences.

The evaluation combines offline model assessment with scenario-based and productivity-oriented analysis. Under the reported evaluation setting, the proposed context-aware wide-and-deep model achieved the strongest predictive performance among the evaluated models, attaining an RMSE of 0.3695 and explaining approximately 81.9% of the variance in observed preference scores. Compared with matrix factorization, neural collaborative filtering, and collaborative filtering baselines, the proposed model reduced prediction error and captured contextual variation more effectively. The broader system-level analysis further indicates that contextual recommendation and Digital Twin-based reasoning can support recommendation relevance, activity planning, decision interpretation, and alignment between user goals and available urban resources. The thesis also identifies important constraints related to data quality, sensing coverage, privacy, scalability, and transferability across urban contexts. Overall, the work shows that personalization in smart cities can be framed and assessed as an integrated, human-centered decision-support problem rather than as an isolated recommendation task.