Integrating Node-Place Model With Shapley Additive Explanation for Metro Ridership Regression

Abstract

Accurate metro ridership estimation is essential for effective urban transportation planning. Traditional regression methods relying on historical passenger flow data often overlook the intrinsic relationship between network scale and land use, limiting predictive accuracy. To address this issue, this study focuses on the Beijing urban rail transit system, aiming to explore how station-level factors influence spatial variations in metro ridership over a decade (2013–2022). An advanced machine learning framework that integrates Node-Place model with Shapley Additive Explanation (SHAP-based machine learning method) is proposed. Twelve influencing factors are developed, including network centrality measures and points of interest (POI) attributes, and a CRITIC weighting method is applied to weight these variables within the Node-Place Model, ensuring an objective assessment of their relative importance. Utilizing LightGBM regression method combined with SHAP values, our model achieved an average R2of 0.88 in annual average daily entry ridership regression and a smaller discrepancy between R2and adjusted R2, significantly outperforming ordinary least squares (OLS, R2= 0.42) and geographically weighted regression (GWR, R2= 0.58). The proposed model consistently demonstrates an AIC value approximately 15% lower than alternative models across multiple years of regression tasks, highlighting its stability and superior performance. Key indicators identified include PageRank and the number of restaurants and enterprises around metro stations. Combining detailed land use data and network centrality measures with advanced machine learning techniques proves to be an effective way to enhancing ridership regression.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 10 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.