Interpretable Machine Learning Reveals Nonlinear Relationships Between Urban Transportation Sustainability and Residents' Quality of Life

Abstract

Urban transportation systems are closely associated with multiple dimensions of urban livability, but conventional additive and linear specifications may provide an incomplete description of nonlinear and interactive patterns among transport-related indicators. This study develops a harmonised analytical panel for 151 metropolitan areas in seven world regions over 2019–2023 (n = 755 city-years; 17 transport-sustainability indicators). The panel combines published city-level anchor statistics with harmonised annual estimates derived from documented regional trends where directly comparable annual city observations were unavailable. A transport-related livability index was constructed from environmental, mobility, safety and economic dimensions and analysed using linear, additive and tree-based models. Gradient-boosted trees were interpreted using Shapley additive explanations, partial-dependence analysis, interaction decomposition and a data-driven breakpoint screening procedure. Under city-grouped cross-validation, LightGBM attained R² = 0.880, compared with R² = 0.826 for ordinary least squares and R² = 0.866 for an additive spline model. Green space per capita, road fatality rate, electric-vehicle fleet share, active travel share and PM2.5 concentration accounted for the largest shares of model attribution. The breakpoint procedure identified candidate changes in the fitted response surface for several mobility-related indicators, including car ownership, transit mode share, active travel and congestion. Model-based scenario analysis further indicated substantial heterogeneity in predicted index changes across policy packages and regions. These results should be interpreted as patterns within a constructed transport-related livability framework rather than as identified causal effects on residents' experienced quality of life. The study demonstrates how interpretable machine learning can be used to examine nonlinear structure, interactions and scenario sensitivity in multi-indicator urban sustainability assessments.