Synthetic Transit Trip-Record Generation Using Deep Generative Models for Travel Demand Modeling
Abstract
Synthetic transit trip-record generation can support activity-based and agent-based travel demand modeling by providing disaggregate records with realistic spatial, temporal, and operational characteristics. Traditional population synthesis methods such as Iterative Proportional Fitting and Iterative Proportional Updating remain widely used, but high-dimensional mixed-type travel data can be difficult to reproduce. This study evaluates synthetic transit trip-record generation, rather than complete household- and person-level population synthesis, using the Minnesota Travel Behavior Inventory 2021 Transit On-Board Survey. Two established deep tabular generative models, Conditional Tabular Generative Adversarial Network (CTGAN) and Tabular Variational Autoencoder (TVAE), are evaluated alongside an independent empirical marginal-sampling baseline. Duration-derived variables are excluded from the primary analysis because 2081 of 2670 records lack observed alighting timestamps. Evaluation combines marginal fidelity, continuous correlation preservation, route–stop and route–time joint distributions, spatial Jensen–Shannon distance, and Train-on-Synthetic, Test-on-Real (TSTR) utility. The independent marginal baseline closely reproduces individual feature distributions but shows substantially weaker correlation preservation and downstream utility, demonstrating that strong univariate fidelity alone does not ensure preservation of task-relevant structure. TVAE shows the strongest continuous correlation fidelity, with a mean absolute correlation difference of 0.037, compared with 0.214 for CTGAN and 0.267 for the independent marginal baseline. However, TVAE has weaker marginal fidelity for boarding-stop ID, alighting-stop ID, and time period than both alternatives. Its TSTR accuracy of 0.530 has an overlapping 95% bootstrap confidence interval with the real-data benchmark accuracy of 0.500, providing no evidence of superior predictive utility. The results demonstrate the importance of complementary marginal, multivariate, spatial, and downstream validation when assessing synthetic transportation data.