Imitate the Right Data: City-wide Mobility Generation with Graph Learning

Realistic simulation of massive human mobility data aids in a range of applications such as traffic management, public transport optimization, and emergency response planning. This task is challenging as most trajectory datasets are sparse. When used for city-level planning, models that merely mimic the sparse input dataset from mobile phones induce errors. In this work, we emphasize the importance of validating simulations with complete trajectories. We present a model for city-wide human mobility generation from sparse data. We first extract spatial, temporal, and activity features from active users' raw trajectory data. With these features, we construct weighted directed graphs to represent individual mobility patterns. We utilize graph convolutional neural networks to learn key parameters in mobility simulation from graphs. With the established parameters, we simulate city-wide mobility with a mechanistic framework. Comparison with baselines demonstrates the efficacy of our method in imitating complete data.