Inferring Behavioral Regimes in Urban Mobility via Spatio-Temporal Optimal Transport

It's eight in the morning on a rainy Tuesday in Manhattan. A Citi Bike dock near Penn Station is nearly empty — all the bikes are gone. Three blocks away, another dock is full. Someone has to decide where to move bikes before the next wave of commuters arrives. That decision depends on one question: where are people going and when? Predicting those origin and destination flows — where trips start and where they end — sounds straightforward, but it isn't. Classical models like the gravity model treat flows as a fixed function of distance and destination attractiveness. They're static. They don't bend to the time of day, and they can't track how behavior shifts when it rains or when it's Saturday. Osipenko and Meng set out to fix that. Their solution borrows from a mathematical field called optimal transport. The core idea is to think of trip origins as supply and trip destinations as demand, then find the minimum-cost plan to move one into the other. Entropic regularization, developed by Cuturi, keeps the solution from collapsing into one extreme routing. You get a smooth distribution of flows across many plausible paths. Osipenko and Meng add a Kullback-Leibler penalty, which measures how far the model's predicted flows diverge from a historical reference. If the penalty weight is large, the model stays close to what riders have done before. If it's small, cost efficiency takes over. That penalty weight — called epsilon — is the conceptual core of the paper. The authors reinterpret it not as a tuning knob but as a behavioral persistence indicator: a dial between habit and efficiency. A high epsilon means riders follow familiar routes. A low epsilon means they're optimizing for cost. Epsilon isn't fixed — it's calibrated day by day across twelve months of Citi Bike data. To confirm that the result isn't fragile, they ran a hundred bootstrap resamples and tried an alternative calibration metric. The two series correlated at 0.99, with tight confidence intervals throughout. The framework also links hourly intervals in a causal chain, so patterns from one hour propagate into the next. Tested on ten neighborhood hubs across the morning peak, the model hits a fit score of around 0.94 on weekdays and 0.91 on weekends. Those numbers hint at the main finding. Weekdays show low epsilon — cost-efficient, stable flows. Weekends show higher epsilon, heavier tails, and more dispersion. Nonparametric tests reject identical distributions at a p-value below 0.01, and a regression confirms a large positive weekend effect with an R-squared of 0.73. Weather adds another layer: precipitation drives epsilon down, consistent with discretionary trips disappearing under bad conditions. Temperature pushes epsilon up, with a significant amplifying effect on weekends. What this means practically is that a static rebalancing schedule built on average flows misses all of this. Monitoring epsilon in real time gives operators a behavioral signal they can act on. The framework was tested on New York City alone, and extending it to other cities or transit modes remains future work. But the core idea holds — a parameter mathematicians treat as a smoothing constant turns out to be a readable signal of how an entire city is choosing to move. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.