Quantifying contributions of geographical features to urban GDP outputs via interpretable machine learning

Abstract

Urban scaling laws, which assume homogeneous population interactions, traditionally describe the relationship between urban population and GDP. However, this approach often overlooks the complexity of urban environments, particularly geographical features such as land use, road networks, and points of interest, which significantly shape urban economies. To address this gap, we propose an interpretable machine learning framework that quantifies the impact of urban geographical features (UGFs) on economic outputs (GDP) across five countries: the USA, Brazil, Nigeria, China, and India. Our study can be summarized in three parts: (1) Using the CatBoost algorithm for GDP estimation, which achieves an average $R^2$ of 0.96 across countries, we demonstrate the substantial effects of UGFs (2) The Shapley Additive Explanations (SHAP) method is employed to quantify feature contributions on GDP, revealing that UGFs account for 45% to 89% variance, with influences differing across and within countries. (3) By classifying cities based on feature contribution vectors, we show that cities with similar GDP levels often exhibit analogous contributions from both population and UGFs, suggesting that shared strategies could be applied to cities with comparable economic profiles. Our findings provide valuable insights into the role of UGFs in shaping GDP, advancing the understanding of how UGFs influence economic development, and offering policymakers more informed suggestions. Furthermore, this framework opens new opportunities to integrate diverse urban features into urban studies through machine learning, enhancing our understanding of the complexity of urban systems.