Nonlinear effects of built environment features on metro ridership: An integrated exploration with machine learning considering spatial heterogeneity

This study explored the nonlinear effects of built environment features on metro ridership and proposes an analytical framework that integrates a gradient boosting decision tree with spatial calibration and validation. Station-level boarding and alighting ridership at different times of the day was obtained from smart card records and used as the dependent variable. Nineteen independent variables, including land use, were calculated based on the directional and size-various catchment area defined by shared bike's origin-destination data. This framework, which accounts for spatial heterogeneity demonstrated strong goodness-of-fit and prediction capability, which has been ignored in previous studies. Furthermore, the proposed framework contributed to modeling based on geographical weighted regression and global machine learning models. Local relative importance mapping of built environment variables revealed varying impacts across Shanghai, diverging from the common practice of averaging into a single value in global machine learning models. Additionally, the nonlinear relationship between influencing variables, such as leisure and shopping, demonstrated a positive trend with boarding and alighting ridership in different periods, and spatio-temporal heterogeneity with the effective range and threshold effect. Rather than focusing on increasing development density to boost metro ridership, this study assesses the saturation of station-level built environment to enable more accurate decision-making based on location, station design, station-area planning, and investment priorities in urban areas.