Coarse-to-Fine Multi-Resolution Diffusion Models for Trajectory Generation in Urban Systems
This paper introduces MR-Traj, a novel multi-resolution diffusion framework that generates large-scale synthetic urban trajectories by modeling them as compositions of coarse-grained milestones and fine-grained segments, thereby improving the capture of complex spatial-temporal patterns, enhancing diversity to reduce privacy linkage risks, and outperforming existing methods in downstream urban mobility tasks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Cities are living, breathing systems, and the most vital pulse they carry is the movement of people. Every day, millions of individuals travel from home to work, from shops to parks, weaving a complex tapestry of routes that defines how a city functions. This data is invaluable for urban planners trying to ease traffic jams, for health officials tracking the spread of illness, and for emergency responders preparing for disasters. Yet, this same data is a treasure trove of private information. Because it reveals exactly where a person lives, works, and spends their time, it is rarely shared openly. To solve this dilemma, scientists have long tried to create synthetic data—computer-generated travel records that mimic the patterns of real life without exposing any single individual. The challenge has always been to make these fake records look real enough to be useful for analysis, while ensuring they are different enough to protect the people they represent.
For years, the best attempts at creating these synthetic journeys focused on getting the big picture right. They ensured that the overall flow of traffic looked similar to the real thing, matching the general density of cars in a city center or the average distance people travel. However, these methods often missed the smaller, finer details. They struggled to reproduce the specific, winding paths a driver might take through a neighborhood or the subtle variations in how people move within a local district. This gap mattered because many real-world tasks, like predicting where a taxi will go next or spotting unusual traffic patterns during a holiday, depend on those local details. If the synthetic data is too smooth or too generic, it fails to support the complex decisions cities need to make.
A team of researchers at the University of Southern California and KDDI Research has introduced a new approach called MR-Traj that addresses this problem by looking at movement on two different scales at once. Instead of trying to generate an entire trip in one go, their system breaks the journey down into a series of major checkpoints, or milestones, and then fills in the detailed path between them. Imagine a traveler planning a trip from one city to another; they first decide on the major stops along the way, and then figure out the specific roads to take between those stops. This new model does something similar. It first uses a computer program to generate a simplified version of a trip, marking only the key turning points and destinations. Then, a second program takes those points and generates the realistic, winding details of the road between them.
The researchers tested this method using massive datasets of real taxi trips from two different cities: Chengdu in China and Porto in Portugal. They compared their new system against several existing methods, including older models based on simple probability and more recent ones using advanced artificial intelligence. The results showed that the new system was just as good as the best existing methods at capturing the overall shape of city-wide movement. However, when it came to the local details, it was significantly better. In tests measuring how closely the synthetic paths matched the real ones in specific neighborhoods, the new model consistently outperformed its competitors. It successfully recreated the intricate, fine-grained patterns of movement that other systems often smoothed over or missed entirely.
Beyond just looking like real data, the synthetic trips generated by this new system proved to be highly useful for practical tasks. When researchers used the fake data to train a computer to predict where a taxi would drop off a passenger, the predictions were nearly as accurate as those made using the real data. The system also excelled at preserving unusual patterns, such as the specific traffic flows that occur during special events or holidays. This is crucial because while normal daily routines are easy to learn, the rare and unusual movements often hold the most important insights for city planners. The ability to capture these anomalies without copying the exact paths of real people suggests the system can support complex urban analysis without compromising the integrity of the data.
Perhaps most importantly, the new method offers a stronger shield for privacy. Because the system generates the major stops and the detailed paths separately, and adds a layer of randomness at each step, the final synthetic trips are more diverse and less predictable than those created by other methods. To test this, the researchers tried to see if they could link a synthetic trip back to the specific taxi that originally generated it. The new system made this task significantly harder for the computer, reducing the success rate of such re-identification attempts compared to other methods. This means the data can be shared more freely for research and planning, with a lower risk that someone could reverse-engineer the information to find out who a specific person is.
The study demonstrates that by thinking about movement in layers—separating the broad structure of a journey from its local details—scientists can create synthetic data that is both highly realistic and safer to use. The researchers found that this two-step process, which they call a coarse-to-fine approach, allows the computer to learn the complex rules of city life more effectively than trying to learn everything at once. While the system is not perfect and still relies on data from specific cities, which may limit how well it applies to every location in the world, it represents a significant step forward. It offers a way to balance the need for detailed, useful information with the urgent need to protect the privacy of the people who make our cities move.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.