Exploring spatiotemporal heterogeneity and dynamics of the built environment impacts on urban traffic congestion with geographically weighted machine learning

Cities are complex multi-scale systems in which transport networks interact dynamically with the built environment, such as population distribution, land use, and transport network structure, leading to traffic congestion patterns that vary across space and time. Considering the complex and dynamic characteristics of traffic congestion, it is essential to explore the spatiotemporal heterogeneity and dynamics in the relationships between built environment factors and urban traffic congestion to develop effective policies that enhances urban livability. Hence, this study employs a geographically weighted machine learning framework that integrates random forest (RF) with geographic weighted regression (GWR), referred to as the geographically weighted random forest (GWRF). Additionally, the SHapley Additive exPlanations (SHAP) method is applied to identify dominant associated factors, interpret nonlinear relationships, and reveal local feature differences between explanatory variables and traffic congestion across different time periods. An empirical case study is conducted in Chongqing, China, a mountainous megacity characterized by complex transport dynamics and strong spatial constraints. The case study utilizes multi-source datasets collected over five months, selects 25 candidate variables that represent built environment characteristics, including land-use diversity, road network design, public transit service, and destination accessibility, as well as demographic and socioeconomic attributes, such as population density and economic indicators. Traffic congestion patterns are examined during morning and evening peak hours on both weekdays and weekends to capture temporal dynamics. Compared with traditional spatial regression models and global machine learning approaches, the geographically weighted machine learning framework achieves about 15-20% higher predictive accuracy. Moreover, the framework exhibits improved stability and adaptability by explicitly incorporating a spatial weighting matrix. From a global perspective, betweenness centrality, office density, bus stop coverage, and shopping density are identified as the dominant factors associated with traffic congestion across the four peak periods. The above results further reveal the nonlinear associations, and threshold effects between key explanatory variables and congestion levels. From a local perspective, the impacts of dominant factors display strong spatial clustering, with the pattern, magnitude, and direction of these associations varying significantly across different spatial regions and time periods. Overall, these findings enhance the understanding of urban transport dynamics, and provide valuable insights for urban planners and operators in developing the planning and management strategies to alleviate traffic congestion and improve urban livability.