TransMode-LLM: Feature-Informed Natural Language Modeling with Domain-Enhanced Prompting for Travel Behavior Modeling

Meijing Zhang, Ying Xu

Keywords: Large Language Models, Travel Behavior Analysis, Mode Choice Prediction, Transportation Planning

Abstract: Understanding traveler behavior and accurately predicting travel mode choice are at the heart of transportation planning and policy making. Traditional methods relying on raw numbers and structured feature representations have limitations on capturing the complex interdependency and qualitative factors that may impact on travel behavior in the real-world, particularly the rich contextual nuances underlying individual decision-making processes. Large language models (LLMs) with promising capabilities for understanding contextual information across domains provides new pathways for travel behavior modeling. In this study, we propose, TransMode-LLM, an innovative framework designed to predict travel modes from natural language descriptions of travelers and their trips. We start by analyzing the importance of features to identify and select key impacting factors (i.e. individual, household and trip characteristics) to enrich context for decision-making. To enhance the performance of LLMs for transportation-specific tasks, we propose a domain-enhanced prompting strategy that incorporates standardize mode definitions. We further explore various learning paradigms (zero-shot and one/few-shot learning) to understand their impact on travel mode prediction using natural language. Finally, we build an evaluation system to compare the performance of the proposed LLM-based approach against state-of-the-art traditional models. Extensive experiments are conducted on the real-world travel survey dataset and the results demonstrate the competitive performance of LLM-based approach such as prediction accuracy compared to the traditional methods. This study advances the application of LLMs in travel behavior modeling, providing promising and valuable insights for both academic research and transportation policy-making in the future.

Submission Number: 20