A comparative machine learning approach to analyze the non-linear dynamics of the built environment in active travel behavior

Understanding how built environment (BE) characteristics influence active travel behavior is critical for promoting sustainable mobility in rapidly urbanizing megacities of the Global South. This study applies a comparative machine learning framework—including logistic regression and seven state-of-the-art ML algorithms—to analyze the non-linear impact of BE attributes on the probability of choosing active travel trips in the Mumbai Metropolitan Region. Among all models, Gradient Boosting Decision Trees (GBDT) demonstrated the highest predictive accuracy (Accuracy = 64%, AUC = 0.70 for work trips), outperforming both traditional and ML-based approaches. Importantly, BE variables proved more influential than socio-demographic factors, accounting for 38% of predictive importance for work trips and 48% for non-work trips. Partial dependence plots (PDPs) revealed clear observed response inflection ranges: active travel likelihood peaks within 200 m of commercial areas, 500 m of bus stops, and in neighborhoods with intersection densities over 400/km 2 . A land-use entropy index between 0.7 and 0.8 was also found to be optimal. These thresholds are directly actionable for urban planners seeking to enhance walkability and reduce car dependence. We present an India-based comparative benchmark that translates non-linear BE effects into probability-scale threshold bands for policy in Mumbai. For cities like Mumbai and similar Global South urban centers, the findings support compact, connected, and mixed-use development strategies that enable more equitable and sustainable mobility systems. The framework is replicable in other high-density cities where comparable travel survey or mobility data are available, enabling locally calibrated built-environment response benchmarks for walkability-oriented planning.