Urban environments generate massive streams of spatio-temporal data, yet accurately predicting urban dynamics remains a fundamental challenge due to complex human mobility patterns, evolving environmental conditions, and distributional shifts across time and space. Mechanism-Driven Explainable Urban Spatio-Temporal Prediction offers a comprehensive and innovative framework that integrates physical mechanisms, causal modeling, and information-theoretic principles into modern deep learning methods, enabling more interpretable, reliable, and generalizable spatio-temporal forecasting. This monograph presents a unified perspective across intrinsic and extrinsic factors that shape urban mobility. It introduces a gravity-inspired potential energy field model to capture intrinsic behavioral mechanisms at both regional and road-network scales, bridging discrete and continuous temporal modeling through differential equation networks. Beyond intrinsic mechanisms, the book proposes a causal basis-vector representation to model spatio-temporal distribution shifts caused by unknown confounders, enhancing robustness under varying scenarios. Furthermore, it develops a theoretically grounded information-theoretic decomposition framework that reduces the complexity of mixed urban data distributions and pushes the predictive performance beyond existing limits. Combining theoretical foundations, methodological innovations, and extensive empirical studies on real-world urban traffic datasets, this book provides a rigorous yet accessible resource for researchers in spatio-temporal modeling, intelligent transportation systems, machine learning, and urban computing. It also serves as a valuable reference for practitioners seeking interpretable and mechanism-aware prediction models for smart city applications.
