A Large-scale Benchmark Dataset for Commuting Origin-destination Matrix Generation
Can Rong, Jingtao Ding, Yan Liu, Yong Li
TL;DR
The paper addresses the challenge of generating reliable commuting origin-destination matrices for regions lacking historical data by introducing a large-scale benchmark dataset covering 3,233 diverse US areas with OD matrices and rich regional attributes. It proposes a new paradigm that treats each area as an attributed directed weighted graph and reframes OD generation as conditional graph edge generation, implemented via a diffusion-based model called WeDAN. Benchmark results show matrix-wise generative models, particularly WeDAN, achieve state-of-the-art performance and better generalization across diverse areas, suggesting a graph-learning approach can capture global mobility patterns beyond single-city transfer. This work enables more generalizable commuting OD generation, with practical impact on urban planning and transportation research by providing a scalable, privacy-preserving data resource and a powerful modeling framework for heterogeneous regions.
Abstract
The commuting origin-destination~(OD) matrix is a critical input for urban planning and transportation, providing crucial information about the population residing in one region and working in another within an interested area. Despite its importance, obtaining and updating the matrix is challenging due to high costs and privacy concerns. This has spurred research into generating commuting OD matrices for areas lacking historical data, utilizing readily available information via computational models. In this regard, existing research is primarily restricted to only a single or few large cities, preventing these models from being applied effectively in other areas with distinct characteristics, particularly in towns and rural areas where such data is urgently needed. To address this, we propose a large-scale dataset comprising commuting OD matrices for 3,233 diverse areas around the U.S. For each area, we provide the commuting OD matrix, combined with regional attributes including demographics and point-of-interests of each region in that area. We believe this comprehensive dataset will facilitate the development of more generalizable commuting OD matrix generation models, which can capture various patterns of distinct areas. Additionally, we use this dataset to benchmark a set of commuting OD generation models, including physical models, element-wise predictive models, and matrix-wise generative models. Surprisingly, we find a new paradigm, which considers the whole area combined with its commuting OD matrix as an attributed directed weighted graph and generates the weighted edges based on the node attributes, can achieve the optimal. This may inspire a new research direction from graph learning in this field.