Highly predictive simple models of human mobility

Understanding how people move between cities and towns is essential for planning transportation systems, designing equitable urban services, predicting disease spread, and tackling environmental challenges, among others. For decades, researchers have used so-called gravity models, which borrow a metaphor from physics: the flow of people between two places grows with the size of their populations and shrinks with the distance between them. These models are simple and interpretable, but often too rough to capture the full complexity of real mobility patterns. At the other end of the spectrum are modern deep learning models that can ingest many data features and predict movement flows with high accuracy, at the price of becoming “black boxes” that offer little insight into the underlying mechanisms driving human mobility.In a new study, ICREA Research Professor Roger Guimerà and collaborators have applied an automated equation-discovery approach known as Bayesian symbolic regression to large mobility datasets. What emerges are closed-form mathematical models that are as accurate—or more so—than the most sophisticated deep learning approaches, yet remain simple and interpretable. These learned models behave like classical gravity models but are tuned directly from data, capturing universal patterns of human mobility across different regions and scales. Strikingly, these models not only match complex algorithms in predictive performance but also generalize better to new geographic areas, suggesting that a few fundamental principles underlie how people choose where to travel. By bridging the gap between interpretability and predictive power, this work offers a new way to model human mobility that could inform smarter infrastructure, better public health responses, and more sustainable urban design.