An Interpretable XGBoost model for Assessing Socio‑Economic Impacts on Predicted and Actual Behavior (Tehran Congestion Pricing Case Study)

Understanding the gap between predicted and actual travel behavior is a critical challenge in the evaluation of congestion pricing policies. This study adopts a machine learning–based approach to model individuals’ travel behavior under congestion pricing using both stated preference (SP) and revealed preference (RP) data. Data were collected from users of Tehran’s central area through SP surveys conducted prior to implementation and RP observations collected after implementation during ۲۰۱۸ and ۲۰۱۹. An XGBoost classification model is employed to capture the complex and nonlinear relationships between socioeconomic characteristics and travel behavior in both SP and RP settings. To ensure transparency and policy relevance, model interpretability is explicitly addressed using SHAP, which quantifies the contribution and direction of each explanatory variable to the model predictions. The SHAP-based analysis enables a systematic comparison of the socioeconomic factors influencing predicted behavior in hypothetical scenarios and actual behavior observed after policy implementation. The results of a comparison of SHAP values between the SP and RP models shows that individuals aged ۴۱–۶۰, women, married individuals, and those who are not household heads are able to predict their travel behavior more accurately when faced with congestion pricing. These findings highlight the value of interpretable machine learning for understanding behavioral discrepancies and improving the design and evaluation of congestion pricing policies.

کلیدواژه ها:

Revealed preference ، Stated preference ، behavior ، XGBoost ، Shap