Explainable AI to understand the impact of life events on travel behavior change
Project description
The XAITraB project develops an explainable and interpretable AI pipeline to analyse how life events—such as employment changes, household relocation, income shifts, or the birth of a child—affect travel behavior over time. Using longitudinal data from the German Mobility Panel, the project examines changes in activity patterns, trip frequencies, distances, and mode choices to better understand behavioral shifts. Machine learning models (random forests, gradient boosting, neural networks) and explainable artificial intelligence (xAI) techniques, such as SHAP values, are applied to uncover the key drivers of mandatory and discretionary activities. Insights from this pipeline are used to inform and improve traditional econometric models, increasing predictive accuracy and enabling more realistic incremental transport modeling. The project serves as a proof of concept toward integrating AI-based insights into next‑generation activity-based models and contributes methodological advances transferable across disciplines.
Results
- SHapley Additive exPlanations (SHAP) analyses identify stability in activity days and highlight key features such as employment status, age, and presence of children.
- Mandatory activity models achieved high predictive performance (pseudo‑R² 77-84%), exceeding those for discretionary activities (41-46%).
- Incorporating machine‑learning‑selected features increased the goodness of fit of traditional econometric models by 5% (mandatory) and 3% (discretionary).
- Top machine learning features were also statistically significant in econometric models, confirming interpretability and robustness.
- Life events such as income changes, birth of a child, and changes in occupation status showed the strongest predictive power across activity types.
Follow-up
Future work aims to integrate AI‑derived insights into incremental activity‑based transport models, enabling year‑to‑year behavioral updating rather than recreating behavior from scratch. Furthermore, a next step could be the development of a generic xAI- and iAI-based methodological pipeline that enables the use of insights from AI-based prediction models in travel demand modeling to achieve a novel state of the art in this domain. This pipeline will inform traditional model estimation for other travel variables, such as traveled distance, mode choice, or vehicle ownership. The results will be used to calibrate an incremental activity-based model. This will lay the foundation for new models for the design and control of future mobility systems.
