GeoFlow: Geospatial-Aware Modeling of Inter-Area Relationships in Origin-Destination Flow Prediction and Generation

Jul 16, 2026

Upload by Ezra

Origin–destination (OD) flow modeling supports urban planning and mobility analysis, yet many graph-based approaches underutilize geographic attributes, weakening their ability to capture long-range and multi-area dependencies. GeoFlow introduces geospatial attribute augmentation with relative positions plus k-hop and geodesic distances, a geometry-intrinsic fusion encoder fusing graph-attention intrinsic signals and coordinate-aware global structure, and an axial-global attention decoder for OD-specific competitive dependencies. Paired with flow matching, GeoFlow generates authentic, diverse mobility samples. Experiments and ablations confirm each component’s value.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

Abstract

Introduction

GeoFlow: Geo-Aware Modeling of Inter-Area Relationships in Origin-Destination Flow Prediction and Generation Zherui Huang 1 Guanjie Zheng 1 Hao Xue 2 3 Linghe Kong 1 arXiv :2607 .05257v 1 [ cs .LG] 6 Jul 2026 Abstract Origin–destination (OD) flow modeling underpins urban planning and mobility analysis, but prevailing graph-based methods often neglect salient geographic attributes, limiting their ability to model long-range and multi-area dependencies. In this paper, we introduce GeoFlow, a novel framework that (i) augments area representations with geospatial attributes, including relative positions, k-hop and geodesic distances, (ii) employs a specialized geometric-intrinsic fusion encoder design that combines graph attention for intrinsic area signals with coordinate-aware encoders for global structure, and (iii) adopts an axial-global attention decoder to capture OD-specific competitive dependencies. For OD flow generation, GeoFlow is paired with flow matching models to produce more authentic and diverse mobility samples. Empirically, GeoFlow achieves superior performance in predictive accuracy, while substantially improving generative fidelity and diversity. Ablation and analytical studies confirm the contribution of each component. Code is available at [https://github.com/ZheruiHuang/GeoFlow](https://github.com/ZheruiHuang/GeoFlow). 1. Introduction Origin–destination (OD) flow describes movements of people or goods between areas, reflecting social and economic activity at a macro level and supporting applications such as urban planning and behavioral analysis (Batty, 2007 ; Zhang et al., 2021 ; Wu et al., 2024) . However, large-scale data collection faces statistical and privacy challenges (Siminiet al., 2021 ; Long et al., 2023), and generalizable models are needed to inform urban development, especially in 1 Shanghai Jiao Tong University 2The Hong Kong University of Science and Technology (Guangzhou) 3The University of New South Wales, Sydney. Correspondence to: Guanjie Zheng <[gjzheng@sjtu.edu.cn](gjzheng@sjtu.edu.cn) >. Proceedings of the 43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026 . Copyright 2026 by the author(s) . data-scarce or early-planning regions (Rong et al., 2025) . Consequently, OD flow prediction and generation have become central research topics (Luca et al., 2021) . Prediction methods estimate flows from area attributes (e.g., sociodemographics), while generation methods further employ random seeds and generative models to capture diverse mobility patterns (Liu et al., 2020 ; Rong et al., 2024) . Despite their conceptual differences, both tasks share common modeling principles, and methods are often evaluated jointly (Rong et al., 2023 ; 2025) . Early approaches, including gravity (Zipf, 1946) and radiation models (Simini et al., 2012), were followed by machine learning techniques such as Random Forest (Breiman, 2001 ; Pourebrahim et al., 2019) and SVR (Drucker et al., 1996 ; Rodr´ıguez-Rueda et al., 2021) . With the rise of deep learning (Dong et al., 2021), models such as DGM (Simini et al., 2021) and GMEL (Liu et al., 2020) achieved notable progress, and recent graphbased methods further advanced prediction and generation by modeling areas as attributed nodes and OD flows as directed weighted edges (Bojchevski et al., 2018 ; Liu et al., 2020 ; Rong et al., 2025) . Although existing state-of-the-art methods represent area networks as graph structures to exploit properties such as translational and rotational equivalent (Liu et al., 2020 ; Satorras et al., 2021), this representation often overlooks easily accessible geographic information, thereby complicating the learning process (Klemmer et al., 2023) . In practice, relationships between areas are encoded in matrix form through pairwise adjacency and straight-line distances (Luo et al., 2024 ; Rong et al., 2025) . While this formulation is theoretically lossless and supports network recons

FAQs

What problem does GeoFlow address in origin–destination (OD) flow modeling?

GeoFlow targets the limitation of graph-based OD methods that neglect geographic attributes, which makes long-range and multi-area dependencies harder to capture.

How does GeoFlow incorporate geographic information into area representations?

GeoFlow augments area representations with geospatial attributes including relative positions, k-hop distances, and geodesic distances, rather than relying on the model to infer these from adjacency matrices.

How does GeoFlow generate OD flows and improve diversity and fidelity?

For generation, GeoFlow is paired with flow matching models to produce mobility samples that are more authentic and diverse, with empirical results showing improved predictive accuracy and generative quality.

Jul 16, 2026

Upload by Ezra