Learning to Deaggregate: Large-scale Trajectory Generation with Spatial Priors

ICLR 2026 Conference Submission24911 Authors

Keywords: trajectory generation, deaggregation, spatial priors, urban mobility, diffusion models

Abstract: Generating realistic large-scale trajectories is essential for applications in urban mobility and transportation, yet current generative models either do not offer any controllability or rely on strong sample-specific conditioning. We introduce the Temporal Deaggregation Diffusion Model (TDDM), a hierarchical framework that first represents mobility using spatial priors, which are marginal distributions over geographical occupancy, and then deaggregates them into trajectories. This separation enables generation without sample-specific conditions, supporting transfer to new regions. To support evaluation, we build a benchmark across three cities spanning different continents (Beijing, Porto, San Francisco), with standardized metrics for fidelity and distributional coverage. Across all datasets, TDDM improves trajectory fidelity and coverage over leading baselines, and demonstrates stable performance when applied to unseen cities. By explicitly decoupling spatial allocation from temporal realization, our work highlights the role of spatial occupancy priors in enabling scalable and generalizable trajectory generation.

Primary Area: generative models

Submission Number: 24911