Examining nonlinearity in population inflow estimation using big data: An empirical comparison of explainable machine learning models

Mobile device location data (MDLD) contain population-representative, fine-grained travel demand information, facilitating opportunities to validate established relations between travel demand and underlying factors from a big data perspective. Using the nationwide census block group (CBG)-level population inflow derived from MDLD as the proxy of travel demand, this study examines its relations with various factors including socioeconomics, demographics, land use, and CBG attributes. A host of tree-based machine learning (ML) models and interpretation techniques (feature importance, partial dependence plot (PDP), accumulated local effect (ALE), SHapley Additive exPlanations (SHAP)) are extensively compared to determine the best model architecture and justify interpretation robustness. Empirical results show that: 1) Boosting trees perform the best among all models, followed by bagging trees, single trees, and linear regressions. (2) Feature importance holds consistently among different tree-based models but is influenced by measures of importance and hyperparameter settings. 3) Pronounced nonlinearities, threshold effects, and interaction effects are observed in relations among population inflow and most of its determinants. 4) Compared with PDP, ALE and SHAP plots are more reliable in the presence of outliers, feature dependency, and local heterogeneity. Taken together, techniques introduced in this study can either be integrated into customary travel demand models to enhance model accuracy or serve as interpretation tools that offer a comprehensive understanding of intricate relations.