Accurate travel behaviour modelling is essential for effective transport planning. Researchers Meijing Zhang and Ying Xu from Singapore University of Technology and Design introduced TransMode-LLM, a cutting-edge framework that integrates statistical analysis with large language models (LLMs) for predicting travel mode choices from survey data. This novel approach enhances traditional methods that often overlook complex contextual factors. Their findings show that domain-enhanced prompting can boost prediction accuracy by up to 42.9%, although efficacy varies by LLM architecture. The framework incorporates a three-phase system: behavioural feature identification through statistics, natural language encoding of structured data, and LLM adaptation via learning paradigms like few-shot learning. Experiments highlighted that models such as o3-mini consistently improve accuracy with limited examples. This research not only provides insights for academic studies but also holds promise for developing data-driven transportation policies, paving the way for a more precise understanding of travel choices.
Source link
Share
Read more