New Diffusion Model Generates Realistic Multi-Resolution Urban Trajectories

Understanding human movement patterns is crucial for various urban applications, including traffic management, epidemic control, and city planning. However, privacy concerns severely limit the availability of large-scale public trajectory data, creating challenges for mobility analysis. Existing methods for generating synthetic trajectories often focus solely on matching global distribution similarities, frequently overlooking the critical mobility patterns that exist at different spatial and temporal resolutions. To address these limitations, a novel multi-resolution diffusion framework, named MR-Traj, has been proposed for generating large-scale synthetic trajectories. MR-Traj explicitly models trajectories by combining coarse-grained milestones with fine-grained segments, allowing it to capture intricate spatial-temporal dependencies across multiple scales. Experimental results demonstrate that MR-Traj not only matches the global distribution similarity of state-of-the-art methods but also significantly surpasses them in accurately modeling fine-resolution mobility patterns and supporting various downstream urban mobility tasks. Furthermore, by introducing stochasticity at multiple resolution levels, MR-Traj generates more diverse trajectories, which empirically reduces the risk of trajectory linkage in seed-guided data releases.