Exploring nonlinear and spatially heterogeneous impacts of built environment on trip generation by geographically weighted random forest and interpretable machine learning

To investigate nonlinear and spatially heterogeneous relationships between the built environment (BE) and trip generation, this study refines the geographically weighted random forest (GWRF) framework by incorporating geographically weighted bootstrapping and weighted training strategies to construct local random forest (RF) models. Accumulated local effect (ALE) plots are further employed to interpret nonlinear patterns. Using Nanjing as a case study, daily aggregated trips are derived from Baidu location-based service data, while BE features are measured using the “7Ds” framework. Results show that the refined GWRF outperforms global and linear benchmark models. Relative feature importance (RFI) indicates that 11 of the 31 variables collectively account for over 95% of predictive importance. Local ALE plots further reveal spatially varying marginal effects, threshold effects, and effective ranges. These findings highlight the locally specific effects of BE on trip generation and can enhance the efficiency and effectiveness of transport planning and policy formulation.