Using machine learning to analyze the factors influencing city-level transportation carbon emissions

Accurately identifying the drivers of transportation-related carbon emissions is essential for supporting city-level low-carbon transitions. This study proposes an interpretable machine learning framework to examine transportation emissions across 298 prefecture-level cities in China from 2000 to 2021. A comprehensive feature set of 203 influencing factors is constructed, covering economic structure, fiscal capacity, land use, infrastructure, and public services. Ten machine learning models are compared based on predictive performance, with CatBoost achieving the highest accuracy and generalization (Test R2: 0.96; Test MAE: 239,716.54). To address the black-box nature of machine learning, SHapley Additive exPlanations (SHAP) are employed to interpret model outputs, revealing nonlinear marginal effects and variable interdependencies. Results show that built-up area, local budgetary revenue, tertiary industry value added, etc. are dominant drivers of transportation emissions, with effect direction and magnitude varying by city context. Compared to previous studies, this research features national-scale coverage, a unified interpretability framework, and clustering analysis for emission typology. SHAP-based clustering identifies three city profiles: (1) High-economy, strong-governance cities with net emission suppression; (2) Transitional cities with mixed and moderate effects; and (3) High-income cities with significant emission growth. These findings underscore the spatial heterogeneity of emission dynamics and the importance of differentiated policy mechanisms—such as tailored emission targets, region-specific infrastructure investment, and adaptive governance strategies. Overall, the study offers a novel framework integrating prediction accuracy, model transparency, and policy relevance, providing actionable insights to guide sustainable urban transport governance in the carbon neutrality era.