With help from AI, your next move can be predicted
Using large language models, Northeastern University researchers can predict human movement. They hope it can help with transportation planning and responding to natural disasters or public crises.
AI might know where you’re going before you do.
Researchers at Northeastern University used large language models, the kind of advanced artificial intelligence normally designed to process and generate language, to predict human movement. RHYTHM, their innovative tool, “can revolutionize the forecasting of human movements,” forecasting “where you’re going to be in the next 30 minutes or the next 25 hours,” said Ryan Wang, an associate professor and vice chair of research in civil and environmental engineering at Northeastern.
The hope is that RHYTHM will improve domains like transportation and traffic planning to make our lives easier, but in extreme cases, RHYTHM could even be deployed to respond to natural disasters, highway accidents and terrorist attacks.
“Accurate prediction is very important for people to understand what’s going on and where they should go,” said Haoyu He, a civil and environmental engineering Ph.D. candidate at Northeastern.
Historically, predicting how people move has been a challenge for one simple reason: Human movements are often random. However, look more closely and it’s possible to see the patterns, or rhythm, of human movement. People might go to the grocery store, school or gym on the same day, at the same time every week.
By using open source mobility data, and the contextual understanding that LLMs have, Wang and He built a model that doesn’t just imitate previous patterns but creates predictions based on certain conditions.
“It’s understanding that people have these 24/7 or weekly or monthly patterns that help to predict people’s location,” Wang said. “It makes the abstract pattern more concrete. That can be interpreted by the model and then used to improve our prediction.”
Editor’s Picks
According to He, RHYTHM is 2.4% more accurate than similar models, but during “irregular periods” where people aren’t following their normal patterns, like weekends, it’s 5% more accurate than the competition. RHYTHM also requires significantly less time to train on data, one of the most time-consuming parts of any LLM’s process.
One of the biggest challenges around mobility prediction has been forecasting where or how one person moves, Wang said. LLMs are already making that easier.
“In the daytime, people living around Boston are going to rush into the city area to work, and then at nighttime they go home, but on an individual level it’s highly random and highly stochastic,” Wang said. “Now with LLMs, that aspect can be incorporated and help us to better predict or represent how this randomness is going to play out.”
Wang and He tested RHYTHM using seven days’ worth of movement data for a group of people and found it could reliably predict movement for the next day, the next few days and the next week. He said that it could theoretically forecast much further into the future, but errors tend to accumulate.
Focusing on that immediate timeframe is actually perfect for where they hope to take RHYTHM next, predicting how people move during natural disasters and other extreme events.
“A lot of times, especially during extreme events, we want to know what happens in the next 24 hours or several hours,” Wang said. “How can we leverage LLMs, which have these more complex characteristics inside of them, to predict those [events] that we hardly observe but are possible to generate or possible to predict even with a very low probability of happening?”
