Spatio-Temporal Forecasting (STF) plays a pivotal role in various smart city applications, such as traffic management, environmental monitoring, and resource allocation. In recent years, deep learning has emerged as the dominant technique for STF due to its powerful representation capabilities. However, existing deep STF models often suffer from scalability and generalization challenges, making them impractical for handling real-world scenarios where spatio-temporal data exhibit partial observability, come at large scale, or are of limited availability.

In this dissertation, we will explore how to bridge these real-world gaps by developing advanced STF models that balance accuracy, scalability, and generalizability. First, from a data perspective, we formulate the semi-supervised STF p...[ Read more ]

Spatio-Temporal Forecasting (STF) plays a pivotal role in various smart city applications, such as traffic management, environmental monitoring, and resource allocation. In recent years, deep learning has emerged as the dominant technique for STF due to its powerful representation capabilities. However, existing deep STF models often suffer from scalability and generalization challenges, making them impractical for handling real-world scenarios where spatio-temporal data exhibit partial observability, come at large scale, or are of limited availability.

In this dissertation, we will explore how to bridge these real-world gaps by developing advanced STF models that balance accuracy, scalability, and generalizability. First, from a data perspective, we formulate the semi-supervised STF problem to facilitate model generalization from monitored regions to unmonitored regions of interest in a city. Building upon this formulation, we propose a self-supervised hierarchical graph neural network, which addresses the semi-supervised STF by extracting transferable knowledge from partially observed spatio-temporal data. Second, from a model perspective, we develop a linear-complexity spatio-temporal graph neural network that efficiently captures the global interactions within the large-scale spatio-temporal system. Finally, from a learning objective perspective, we introduce a lightweight multi-domain pre-training approach to achieve few-shot and zero-shot forecasting across diverse domains and spatial regions.

This approach enhances the practical usability of STF models in data-scarce scenarios, while the lightweight nature facilitates rapid model deployment in resource-constrained urban environments. We envision that the proposed methodologies will pave the way for the widespread application and deployment of STF models in real-world settings.