Learning Traveler-Specific Substitution Structures for Emerging Mobility Choice: A Persona-Conditioned Graph Neural Network

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Emerging mobility services such as Urban Air Mobility (UAM) complicate mode-choice prediction because travelers may position them differently relative to existing modes. Conventional econometric and machine-learning models mainly capture substitution indirectly through error distributions or population-level structures. They are not designed to flexibly capture traveler-specific pairwise substitution patterns across multiple modes. This study proposes a Persona-Conditioned Graph Neural Network (PC-GNN) that learns individual substitution structures with nonlinear preferences. Because such structures are difficult to identify from sparse stated-choice observations alone, large language models (LLMs) are introduced to elicit diverse priors from a subset of travelers and store them in a reusable belief bank. The priors contain persona vectors and substitutability matrices, representing the individual substitution structures of these travelers. For the remaining travelers, observed attributes are combined with retrieved LLM-elicited priors to construct multiple conditioned GNNs, and a retrieval-mixture module then produces the final prediction. We further relate PC-GNN to established discrete choice models and derive an approximation bound to characterize this retrieval-mixture mechanism. Using a stated-preference survey of nearly 1,600 respondents in the Guangdong-Hong Kong-Macao Greater Bay Area, the analysis reveals heterogeneity in UAM substitution. High-speed rail (HSR) emerges as its primary substitute, while car-oriented travelers exhibit a substitution pattern involving both HSR and private cars. PC-GNN also improves the mode-choice predictive accuracy by about 13% over second-best baselines. Ablation and scalability analyses further confirm the value of persona augmentation, individual-specific graphs, and behaviorally informed retrieval.

This publication has 44 references indexed in Scilit: