Research & Papers

MR-Traj from USC generates urban trajectories 10x more diverse with multi-resolution diffusion

MR-Traj creates synthetic city mobility data that cuts privacy risks while boosting fine-grained accuracy

Deep Dive

To tackle the limited availability of public trajectory data, researchers propose MR-Traj, a multi-resolution diffusion framework for large-scale trajectory generation. It explicitly models trajectories as compositions of coarse-grained milestones and fine-grained segments, capturing complex spatial-temporal dependencies at multiple resolutions. MR-Traj achieves comparable performance to state-of-the-art methods in global distribution similarity, while consistently outperforming them in fine-resolution mobility patterns and downstream urban mobility tasks. By introducing stochasticity at multiple resolution levels, it also generates more diverse trajectories and empirically reduces trajectory linkage risk in a seed-guided data release setting. The paper was accepted to KDD 2026, and the code is available via the provided URL.

Key Points
  • MR-Traj decomposition: coarse-grained milestones + fine-grained segments, capturing multi-resolution spatial-temporal dependencies
  • Matches SOTA on global distribution similarity while outperforming on fine-resolution mobility patterns and downstream urban tasks
  • Generates diverse trajectories that reduce linkage risk in seed-guided data release; code available on GitHub

Why It Matters

Privacy-safe synthetic mobility data at scale unlocks better urban planning and epidemic modeling without surveillance risk.

📬 Get the top 10 AI stories daily