Yuan Yuan, Ding Jingtao, Jin Depeng, Li Yong
Beijing National Research Center for Information Science and Technology (BNRist), Department of Electronic Engineering, Tsinghua University, Beijing 100084, P. R. China.
PNAS Nexus. 2025 May 6;4(5):pgaf081. doi: 10.1093/pnasnexus/pgaf081. eCollection 2025 May.
City-scale individual movements, population flows, and urban morphology are intricately intertwined, collectively contributing to the complexity of urban mobility and impacting critical aspects of a city, from socioeconomic exchanges to epidemic transmission. Existing models, derived from fundamental laws of human mobility, often capture only partial facets of this complexity. This article introduces DeepMobility, a powerful deep generative collaboration network designed to encapsulate the multifaceted nature of complex urban mobility within one unified model, bridging the gap between the heterogeneous behaviors of individuals and the collective behaviors emerging from the entire population. As the first generative deep learning model to integrate micro- and macrolevel dynamics through bidirectional collaboration, DeepMobility generates high-fidelity synthetic mobility data, overcoming key limitations of prior approaches. Our experiments, conducted on mobility trajectories and flows in cities of China and Senegal, reveal that unlike state-of-the-art deep learning models that tend to "memorize" observed data, DeepMobility excels in learning the intricate data distribution and successfully reproduces the existing universal scaling laws that characterize human mobility behaviors at both individual and population levels. DeepMobility also exhibits robust generalization capabilities, enabling it to generate realistic trajectories and flows for cities lacking corresponding training data. Our approach underscores the feasibility of employing generative deep learning to model the underlying mechanism of human mobility and establishes a versatile framework for mobility data generation that supports sustainable and livable cities.
城市尺度的个体移动、人口流动和城市形态相互交织，共同导致了城市交通的复杂性，并影响着城市的诸多关键方面，从社会经济交流到疫情传播。现有的基于人类移动基本规律的模型，往往只能捕捉到这种复杂性的部分方面。本文介绍了深度移动模型（DeepMobility），这是一个强大的深度生成协作网络，旨在将复杂城市交通的多面性封装在一个统一模型中，弥合个体异质性行为与整个人口出现的集体行为之间的差距。作为首个通过双向协作整合微观和宏观层面动态的生成式深度学习模型，深度移动模型生成高保真的合成交通数据，克服了先前方法的关键局限性。我们在中国和塞内加尔城市的移动轨迹和流量上进行的实验表明，与倾向于“记忆”观测数据的现有先进深度学习模型不同，深度移动模型擅长学习复杂的数据分布，并成功再现了表征个体和人口层面人类移动行为的现有通用标度律。深度移动模型还表现出强大的泛化能力，使其能够为缺乏相应训练数据的城市生成逼真的轨迹和流量。我们的方法强调了采用生成式深度学习对人类移动潜在机制进行建模的可行性，并建立了一个支持可持续和宜居城市的通用交通数据生成框架。
