Effective Urban Region Representation Learning Using Heterogeneous Urban Graph Attention Network (HUGAT)

Background: Understanding the complex dynamics of urban environments is crucial for building smarter and more livable cities. To achieve this, it is essential to capture the interactions between physical space and human activities at finer scales. Objective: This study aims to develop a model that effectively represents urban regions by incorporating both spatial features and human activity patterns, in order to better understand and predict urban dynamics. Methods: We propose a novel urban region representation model called the Heterogeneous Urban Graph Attention Network (HUGAT). HUGAT utilizes an urban-Heterogeneous Information Network (Urban-HIN) to model diverse relations among multiple urban node and edge types. It simultaneously learns multiple objectives of spatial and human activity variations through a heterogeneous graph attention network. Results: Experiments conducted on data from New York City show that HUGAT outperforms state-of-the-art models across various prediction tasks, including average personal income, poverty ratio, region popularity, and spatial clustering. The results demonstrate not only the effectiveness of HUGAT but also the importance of human movement data in capturing ever-evolving urban dynamics. Conclusion: HUGAT’s ability to utilize readily available urban data sources has the strong potential to enhance informed decisions on urban policy and public engagement by providing critical insights in a timely and cost-effective manner, without relying on direct information sourcing such as surveys.