SEDAN, a diffusion model for cross-city commuting-flow prediction, achieves 7.38% RMSE improvement over prior baseline by fusing urban semantics and spatial structure

Researchers propose SEDAN, a Structure-Enhanced Diffusion model that represents cities as attributed graphs, where regions are nodes with demographic and point-of-interest features, and commuting flows are weighted edges.

The model combines semantic information (regional attributes processed through graph-transformer node interactions) with spatial structure (adjacency and distance matrices that guide attention and serve as diffusion conditions) to generate origin-destination (OD) matrices—mathematical representations of commuting patterns between city regions.

Experiments on U.S. city datasets show SEDAN achieves a 7.38% improvement in RMSE compared with the state-of-the-art baseline WEDAN, and remains robust across heterogeneous urban scenarios and varying structural patterns.