Learning to generate synthetic human mobility data: A physics-regularized Gaussian process approach based on multiple kernel learning

Passively-generated mobile data has grown increasingly popular in the travel behavior (or human mobility) literature. A relatively untapped potential for passively-generated mobile data is synthetic population generation, which is the basis for any large-scale simulations for purposes ranging from state monitoring, policy evaluation, and digital twins. And yet, this significant potential may be hindered by the growing sparsity or rate of missingness in the data, which stems from heightened privacy concerns among both data vendors and consumers (users of service platforms generating individual mobile data). To both fulfill the great potential and to address sparsity in the data, there is a need to develop a flexible and scalable model that can capture individual heterogeneity and adapt to changes in mobility patterns. We propose a conditional-generative Gaussian process framework that learns kernel structures characterizing individual mobile data and can provably replicate observed patterns. Our approach integrates physical knowledge to regularize the framework such that the generated data obeys constraints imposed by the built and natural environments (such as those on velocity and bearing). To capture travel behavior heterogeneity at the individual level, we propose a data-driven multiple kernel learning approach to determine the optimal composite kernel for every user. Our experiments demonstrate that: (1) the impact of kernel choice on mobility metrics derived from synthetic data is non-negligible; (2) physics-regularization not only reduces model bias but also improves uncertainty estimates associated with the predicted locations; and (3) the proposed method is robust and generalizes well to varying individuals and modes of travel.