Fusing Urban Structure and Semantics: A Conditional Diffusion Model for Cross-City OD Matrix Generation

2026-05-05 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Public Transportation & Urban Mobility · Depth: Expert, extended

Summary

SEDAN, a Structure-Enhanced Diffusion model conditioned on Attributed Nodes, is proposed for generating cross-city Origin-Destination (OD) matrices. This model addresses the challenge of considerable heterogeneity in commuting patterns across cities by fusing urban structure and semantics. SEDAN models a city as an attributed graph where regions are nodes with demographic and point-of-interest features, and commuting flows are weighted edges. It incorporates adjacency and distance matrices to characterize spatial structure, designing a fusion mechanism to jointly model semantic information and spatial constraints. Experiments on the LargeCommuingOD dataset, covering 3,233 U.S. spatial units, show SEDAN improves RMSE by 7.38% and significantly reduces JSD for inflow, outflow, and total OD flow by 10.34%, 22.22%, and 26.67% respectively, compared to the WEDAN baseline. The model demonstrates robustness across heterogeneous urban scenarios and varying structural patterns, providing a generalizable solution for urban governance and traffic planning.

Key takeaway

For urban planners and transportation analysts tasked with generating accurate commuting flow predictions across diverse cities, SEDAN offers a robust solution. Its explicit fusion of urban semantics and spatial structures significantly improves the accuracy and structural consistency of generated OD matrices, especially in complex polycentric environments. You should consider integrating such structure-enhanced diffusion models to overcome data scarcity and heterogeneity challenges, leading to more reliable insights for urban planning and resource allocation.

Key insights

Fusing urban semantics and spatial constraints via a conditional diffusion model enhances cross-city OD matrix generation accuracy and generalizability.

Principles

• Explicitly fuse semantic and spatial information for robust OD modeling.
• Adjacency and distance priors provide complementary local and global constraints.
• Logarithmic transformation stabilizes training for long-tail flow distributions.

Method

SEDAN models cities as attributed graphs, using a spatially-guided conditional diffusion framework. It integrates regional attributes (demographics, POIs) and spatial priors (adjacency, distance matrices) into a graph transformer-based denoising network to predict edge-level noise.

In practice

• Use graph transformers to capture nonlinear regional interactions.
• Apply DDIM for accelerated inference in diffusion models.
• Incorporate demographic and POI data for latent travel demand.

Topics

• OD Matrix Generation
• Conditional Diffusion Models
• Graph Transformers
• Urban Spatial Structure
• Cross-City Generalization

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Understanding Long-Term Dynamics of Individual Metro Usage: A Hidden Semi-Markov State Framework with Survival Analysis — The framework models individual metro usage lifecycles by integrating latent mobility states with exit and re-entry hazards.
• EditSSC: Toward Editable Semantic Occupancy Scenes with Unconditional Diffusion Models — EditSSC repurposes 2D latent diffusion for 3D semantic scene generation, enabling training-free editing capabilities.
• TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation — End-to-end, map-free transit route generation is feasible by training LLMs on large-scale trajectory data, enabling implicit spatial grounding.
• CityTrajBench: A Unified Benchmark for City-Scale Vehicle Trajectory Generation — CityTrajBench unifies urban trajectory generation evaluation, revealing multi-objective quality and no single model dominance.
• PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain — Improved data quality and multi-token latent spaces enable stable, high-performance, low-latency 3D scene completion.
• Generative Modeling on Metric Graphs via Neural Optimal Transport — A neural optimal transport framework enables continuous generative modeling on metric graphs by embedding, solving, and projecting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.