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

Trajectories arise in many settings, including urban mobility, transportation, and maritime traffic. Existing generative models either offer no control over output distributions or rely on conditioning information specific to each individual trajectory, such as its origin and destination, distance, or departure time. This sample-specific conditioning limits controllability and ties the model to the environment where those statistics were observed. We propose to separate _where_ movement occurs from _how_ it unfolds: regional movement is summarized by a spatial prior $\Pi$, a marginal distribution of occupancy aggregated over many trajectories, and the generative model is trained so that samples drawn conditionally on $\Pi$ aggregate to match it. We call this property _marginal consistency_, which turns the spatial prior into a controllable input, enabling zero-shot generation in unseen regions, cities, and even unseen domains by supplying only the target region's prior. The Temporal Deaggregation Diffusion Model (TDDM) realizes this idea. We evaluate it across four datasets spanning three continents and three modalities: multi-modal human mobility (Geolife), urban taxi (Porto, Cabspotting), and maritime traffic (Brest). TDDM achieves improved fidelity and coverage over leading baselines and stable performance when transferred across regions, cities, and domains.