Urban Region Representation Learning: A Positional and Structural Graph Approach
Urban environments are composed of intricate spatial structures and movement patterns, impacting various urban elements like transportation, land use, and economic activity. Firstly, in this dissertation, we review the geospatial artifical intelligence literature on urban region representation models, categorizing them based on learning paradigms, spatial data modalities, and architectural design. Furthermore, we highlight key challenges on region representation models such as the data quality issue, which some datasets are incomplete, presenting mising labeled data. Additionally, urban region models overlook spatial heterophily, where adjacent regions can have varied functions, being an intrinsic characteristic of geographic data. To resolve this, we present HAVANA (Hybrid Attentional Graph Convolutional Network for Semantic Venue Annotation), a model using Graph Neural Networks and human mobility data to advance POI classification improving geographic data quality. Building upon this improved data representation, we introduce a novel spatial heterophily-aware Graph Transformer named FisherGT which is incorporated on HAMURE (Heterophily-Aware Urban Multi-View-Based Region Embedding), a self-supervised multistage model. Hence, HAMURE improves representation quality, yielding better outcomes for land use clustering, crime prediction, population density estimation.