Graph Learning for Urban Computing: Techniques and Applications

Abstract

Urban systems are relational, heterogeneous, multi-scale, and shaped by operational constraints. Reliable urban decision support therefore requires methods that model relationships among urban entities, capture dependencies across scales, and produce solutions that remain feasible in practice. This dissertation develops graph learning frameworks for four urban computing tasks: sensor recovery, spatial generation, service boundary redesign, and infrastructure expansion. The first two are state-oriented, inferring or generating urban conditions within graph-based systems; the latter two are intervention-oriented, reshaping the relational structures that organize services and infrastructure. For sensor recovery, this work develops an adaptive graph convolutional imputation framework that reconstructs missing environmental observations by learning latent inter-sensor dependencies and integrating them with bidirectional temporal modeling. For spatial generation, it formulates instruction-guided land use planning as conditional graph generation and introduces a hierarchical graph generative adversarial framework that captures global planning structure before refines detailed configurations through local spatial regularity and long-range functional dependencies. For service boundary redesign, it frames school redistricting as graph partition editing under contiguity and capacity constraints, using a sampling-based framework that produces feasible alternatives while surfacing trade-offs among enrollment balance, spatial compactness, and stakeholder impact. For infrastructure expansion, it models transit network growth as constrained sequential subgraph expansion and presents a feasibility-aware graph reinforcement learning framework that learns expansion policies under budget and engineering constraints. Across these frameworks, graph structure bridges state estimation and structural intervention by capturing latent dependencies for recovery and generation while defining the feasible decision space for redesign and expansion. This dissertation further identifies limitations in graph-based urban decision support, including representation quality, scalability, objective alignment, and institutional usability. The studies demonstrate that graph learning is most effective for urban decision support when graph structure is treated not only as a representation of urban complexity, but also as a mechanism for adaptive estimation, generative reasoning, and feasible intervention under real-world planning constraints.