AI just got better at inventing fake commuters, and city planners should care.
A new research paper describes a two-stage generative framework for building synthetic populations, the fabricated but statistically plausible datasets that feed activity-based travel demand models. Stage one uses a Wasserstein GAN with gradient penalty, augmented with three regularization terms called IGP, LDR, and CLAP, to produce tabular profiles that are more feasible, diverse, and novel than a standard WGAN-GP. Stage two runs Transformer and LSTM-Attention models on top of those profiles to generate sequential travel details like departure time, trip purpose, and travel mode. Against the baseline, the regularized approach lifted feasibility by 2.1 to 3.7 percentage points, novelty by 6.6 to 10.0 points, and F1 score by 6.3 to 8.6 points; for sequence generation, the Transformer model hit 90.6 percent F1 versus 89.1 percent for LSTM-Attention, though LSTM-Attention matched real-world trip-length patterns more closely.
The point isn't a flashier model. Travel demand forecasts, the numbers that justify new highways, transit lines, and zoning decisions, are only as good as the synthetic populations built from thin survey data. When surveys miss valid combinations of people and trips, known as sampling zeros, models either ignore real-world diversity or hallucinate impossible combinations, known as structural zeros. Getting that balance right is unglamorous, but it's the kind of infrastructure work that quietly shapes which transit line gets funded.
No chatbot demo here, just better plumbing. That's what most AI research actually looks like once you get past the funding announcements.