Reproduce “Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion” on OpenResearch

About this paper

Baoshen Guo, Zhiqing Hong, Junyi Li, Shenhao Wang, Jinhua Zhao

Researchers from Singapore-MIT Alliance for Research and Technology, Rutgers University, University of Florida, and MIT developed Cardiff, a cascaded hybrid diffusion model that generates high-fidelity, fine-grained urban mobility trajectories by leveraging spatial hierarchy. This two-stage approach, which first produces coarse road segments and then refines them into GPS points, achieved lower Jensen-Shannon Divergence scores (e.g., JSD-SD of 0.0043 on the Singapore dataset) than baseline methods and offered flexible privacy-utility trade-offs.

Key insights

• Leveraging the inherent spatial hierarchy of urban mobility data (coarse road segments and fine GPS points) within a cascaded generative framework effectively addresses the complexity of trajectory synthesis.
• Guiding fine-grained GPS generation with coarse-grained road segment information ensures spatial consistency and adherence to the underlying road network, enhancing the realism of synthetic trajectories.
• Encoding discrete road segments into a low-dimensional latent space significantly improves the efficiency and scalability of the diffusion process, making fine-grained trajectory synthesis practical for large-scale datasets.

Summary by alphaXiv.