Nonlinear responses of station-area functional vitality to the built environment in urban rail transit networks: an interpretable machine learning approach

Abstract

Functional vitality in urban rail transit station areas reflects the concentration of services and urban functions around stations. This study examines 170 operational stations on Lines 1–8 of the Ningbo rail transit network. Using point-of-interest (POI), bus stop, street network, station attribute, and rail network data within 800 m catchments, we constructed a station-area functional vitality index and investigated its built-environment determinants using XGBoost, SHapley Additive exPlanations (SHAP), and partial dependence plots (PDPs). XGBoost outperformed the comparison models, with an R 2 value of 0.7213, a root mean square error (RMSE) of 0.0801, and a mean absolute error (MAE) of 0.0589. Distance to the city center, intersection density, and the number of bus stops were the dominant predictors, together accounting for 71.55% of total importance. One-dimensional PDPs revealed marked nonlinear responses: predicted vitality increased at approximately 105–113 intersections/km 2 and around 16–17 bus stops but declined rapidly as distance from the city center increased within the first 4 km. Two-dimensional PDPs further suggested that central location conditions the effects of other built-environment factors, while strong street connectivity combined with sufficient bus stop provision corresponds to higher predicted vitality. These findings provide evidence for station-area functional planning and bus–rail integration.