Agent-Based Simulation of Human Mobility Using High-Resolution Foot-Traffic Data

Abstract

Accurate models for human mobility are crucial in simulations of disease transmission, autonomous transportation, and urban planning. However, contemporary mobility models rely on random representations of human mobility due to insufficient data sources. We compare the accuracy of human mobility using three distinct statistical models: a probabilistic and Latent Dirichlet Allocation (LDA) model that are calibrated using foot-traffic data from SafeGraph and a random model that delineates human mobility without being data-driven. We fit the probabilistic and LDA models on a single month and evaluate their respective accuracies on empirical data from subsequent months using the Jaccard similarity coefficient. Our findings demonstrate that the probability model produces the most accurate inferences of human mobility, followed by the LDA model, with the random model yielding the lowest accuracy. Our future steps include training the probability and LDA models on a stratified population of location visits to account for potential overfitting.