Exploring the Impact Mechanisms of the Built Environment on Bike-Sharing Usage: A Study Based on Random Forest and Multiscale Geographically Weighted Regression

Bike-sharing is an important component of urban short-distance mobility systems, offering a flexible, low-carbon alternative to motorized travel while addressing “last-mile” connectivity in metropolitan areas. Despite its rapid expansion, bike-sharing usage exhibits pronounced spatiotemporal variability and spatial heterogeneity, which complicates system management, infrastructure provision, and policy design. Studies have predominantly relied on global statistical models or localized spatial methods, making it difficult to simultaneously identify the dominant influencing factors and reveal their spatially varying effects. To address this limitation, this study proposes an integrated analytical framework combining Random Forest (RF) with Multiscale Geographically Weighted Regression (MGWR) to explore the mechanisms through which the built environment influences bike-sharing usage from global and local perspectives. Using large-scale bike-sharing trip data from central urban Shenzhen, China, a comprehensive set of built environment indicators was constructed, covering the road network structure, population and building density, land-use characteristics, points of interest, and public transport accessibility. First, RF was employed to capture complex nonlinear relationships between bike-sharing usage and the built environment, and SHapley Additive exPlanations (SHAP) values were used to quantify the global importance and marginal effects of each variable. Based on the RF-SHAP results, road network, population, work–life POI, and metro station densities within a 1 km buffer were identified as the most influential predictors of bike-sharing demand. Subsequently, these key variables were incorporated into the MGWR model to examine their spatially varying effects and scale-dependent influence mechanisms. The results revealed substantial differences in the magnitude and spatial scale of the built environment effects. Road network density exhibited strong local-scale effects, indicating pronounced spatial heterogeneity that is closely associated with micro-level street connectivity and cycling infrastructure conditions. Its positive influence on bike-sharing demand was notable in major commuting corridors and high-connectivity areas. Conversely, population density demonstrated near-global-scale effects with relatively stable coefficients across space, suggesting that overall demand intensity provides a consistent baseline for bike-sharing usage throughout the study area. Work-life POI density showed context-dependent effects, with negative associations observed in highly mixed-use urban cores, particularly on non-workdays. This pattern suggests a substitution effect in compact, functionally intensive areas where walking may replace short-distance cycling. Metro station density within a 1-km buffer exerted a robust and spatially stable positive influence on bike-sharing usage, highlighting the importance of bike-sharing as a feeder mode for rail transit. The effect was pronounced in areas with weaker bus service coverage, underscoring the importance of “metro-bike” integration in enhancing multimodal accessibility. Overall, built environment factors influenced bike-sharing usage through distinct spatial scales and heterogeneous mechanisms that cannot be fully captured by single-model approaches. By integrating global importance identification with local-scale spatial heterogeneity analysis, the proposed RF-MGWR framework provides a more comprehensive understanding of bike-sharing behaviors. This study combines machine learning and spatial econometric methods and offers empirical evidence on the scale-dependent effects of the built environment on active travel. These findings are important for differentiated bike-sharing management, cycling infrastructure investment, and integrated public transport planning, thereby supporting more efficient and sustainable urban mobility strategies.