Practical perspectives on machine learning interpretability for public transport demand forecasting

Galiger, Gergő and von Münster, Max and Nyerges, Márk and Tóth, Patrik and Kovács, Péter (2026) Practical perspectives on machine learning interpretability for public transport demand forecasting. In: Re-Generation in Transport, Proceedings of the 11th TRA Conference. Lecture Notes in Mobility . Springer, Budapest. ISBN 978-3-032-37192-8 (In Press)

Preview

Text E3.pdf - Accepted Version Download (437kB) | Preview

Abstract

As the data generated in public transport systems is growing in complexity and quantity, data-driven artificial intelligence models are increasingly used to optimize operations, improve passenger experience, and support decision-making. However, the adoption of machine learning (ML) in this domain is impeded by the lack of interpretability of these black-box techniques. While often highly accurate, these models struggle to gain the trust of transport authorities and stakeholders due to their low transparency. This paper explores the role of interpretable ML in public transport demand forecasting based on a real-world dataset from the transport authority of Budapest. In particular, we perform a fair comparison of the inherently interpretable NeuralProphet time series model, with the post-hoc interpretation of XGBoost using SHAP on data from the metro line M4 of the city. Our results demonstrate that XGBoost outperforms NeuralProphet in terms of forecasting accuracy. Additionally, we show that the insights from the post-hoc SHAP interpretation aligns with both the seasonality components of NeuralProphet and domain knowledge from transportation experts. These results support the adequacy of using SHAP to interpret accurate models for predicting public transport demand, with implications in the broader domain of timeseries forecasting. Source code is available at: github.com/galigergergo/PTDemComp.

Actions (login required)

View Item