Learning the complexity of urban mobility with deep generative network

Area of Science:

• Urban mobility dynamics
• Deep learning applications
• Socioeconomic and public health modeling

Background:

• Urban mobility is complex, influenced by individual movements, population flows, and urban form.
• Existing models often capture only limited aspects of urban mobility.
• Bridging micro- and macro-level dynamics is crucial for comprehensive understanding.

Purpose of the Study:

• To introduce DeepMobility, a deep generative collaboration network for unified urban mobility modeling.
• To generate high-fidelity synthetic mobility data by integrating heterogeneous individual and collective behaviors.
• To overcome limitations of existing models in capturing multifaceted urban mobility.

Main Methods:

• Developed DeepMobility, a novel deep generative collaboration network.
• Integrated micro- and macro-level dynamics through bidirectional collaboration.
• Validated on mobility data from cities in China and Senegal.

Main Results:

• DeepMobility successfully learns intricate data distributions, unlike models that memorize data.
• Reproduces universal scaling laws of human mobility at individual and population levels.
• Demonstrates robust generalization, generating realistic data for cities without training data.

Conclusions:

• Generative deep learning is feasible for modeling human mobility mechanisms.
• DeepMobility provides a versatile framework for generating urban mobility data.
• The approach supports the development of sustainable and livable cities.