Exploring the Potential of Large Language Models in Daily Travel Activity Pattern Prediction

Abstract

Daily Travel Activity Patterns (DAPs/DTAPs) are critical in transportation modelling, especially within Activity-Based frameworks, enabling planners and policymakers optimize network performance and operational efficiency. While multi-label classification approaches have been previously applied for DAP prediction, the comparison of Large Language Models (LLMs) with other widely used algorithms in this domain remains unexplored. Leveraging the strengths of LLMs in sequence modelling and multi-label classification, this paper makes a novel contribution by implementing and comparing several open-access LLMs for DAP prediction against traditional discrete choice models (Multinomial Logit) and popular machine learning and deep learning algorithms. Using the 2018 Origin-Destination Montreal travel survey data, including socio-demographic information on over 169,000 individuals, results demonstrate that LLMs can improve prediction accuracy by up to 3% over other methods. However, this improvement comes with significantly higher training times, highlighting a trade-off between accuracy and computational efficiency.