Heterogeneous nonlinear effects of metro network structure on station-level ridership: evidence from varying time periods, flow directions, and station types

Understanding station-level ridership is essential for metro planning, yet existing studies neglect the role of network structure and its heterogeneous nonlinear effects. This study fills this gap by integrating complex network theory with the XGBoost‑SHAP method. Four network structure measures (NSMs) are employed to characterize network structure: degree centrality, eigenvector centrality, harmonic closeness centrality, and mean travel-time weighted degree. Using Beijing metro data, we categorize stations into three types—A, B, and C—based on temporal ridership patterns, facilitating analysis across 12 modeling contexts incorporating two time periods, two flow directions, and three station types. Results show that NSMs are important influencing factors of station-level ridership. Their overall contribution ranges from 16.6% to 47.9% across the 12 scenarios, ranking among the top two variable categories. In addition, the four NSMs exhibit pronounced nonlinear effects on station-level ridership, showing notable heterogeneity across modeling contexts.