Would you like to know how crowded your drive to the beach will be in three weeks? Or where your ex will be on a Friday night next month so that you can avoid him?
Adam Sadilek, formerly of Microsoft, and John Krumm, a principal researcher at Microsoft, were inspired by the question of predicting where people would be in the future and even led off with the query, “Where are you going to be 285 days from now at 2PM?” in their paper, Far Out: Predicting Long-Term Human Mobility.
“At first glance,” the researchers told us, “it sounds like a very difficult problem.”
Sadilek, Krumm, and others have done a lot of research on predicting where a person might be in the immediate future--say, in an hour or two. Logically enough, it's been found that a person’s previous location is a good clue for their next location. But as these models are extended into the future, they give poorer and poorer results. To guess with any accuracy where someone would be in 20 or 200 days would be more of a challenge. In order to do so, Sadilek and Krumm realized, they’d have to develop new techniques.
Using information from a pool of 300 volunteers in the Seattle metro area, Sadilek and Krumm gathered a mountain of location data. As the volunteers went about their daily lives--going to work, to the grocery store, out for a jog, even for transcontinental travel--each carried a GPS device much the same way they carried a cell phone. To further ensure accuracy, the researchers also installed GPS devices in commercial shuttles and transit vans that the volunteers used regularly, and the volunteers’ own vehicles. After collecting over 150 million location points, the researchers then had Far Out, the first system of its kind to predict long-term human mobility in a unified way, parse the data. Far Out didn't even need to be told exactly what to look for--it automatically discovered regularities in the data.
“For example, it might notice that Tuesdays and Thursdays are usually about the same and fairly consistent from week to week,” the researchers told us. “Then when we ask about a future Tuesday or Thursday, the algorithm automatically produces a typical Tuesday/Thursday as a prediction.”
Salidek and Krumm were pleasantly surprised with the results. It turns out that no matter how spontaneous we think we are, humans are actually quite predictable in our movements, even over extended periods of time. Not only did Far Out predict with high accuracy the correct location of a wide variety of individuals, but it did so even years into the future.
When we asked how Far Out compensates for people who change jobs, cities, spouses or even just neighborhoods, the researchers said that the Far Out model adapts to new data. “If there is a sharp transition, such as a move to another city, the system notices there is a discrepancy between its predictions and actual data and adapts to the new patterns,” the researchers said. “Most people have only a few ‘revolutionary’ changes in the course of their lives, so Far Out isn't caught off guard too often.”
For now Far Out is strictly a research project not yet available in commercial products or services. And although its focus currently is on the future whereabouts of single individuals, eventually, the researchers' hope is that it can be applied to larger populations. This could be a boon to urban planners by leading to more accurate predictions about the spread of disease, traffic congestion, and the demand for electricity.
Marketers and advertisers, too, would relish the opportunity to target our future selves with ads like, “Need a haircut? In four days, you’ll be 100 yards from a salon that will have a $15 special.” On the social side, there could even be something like a Foursquare of the Future--who wouldn’t want to know where their friends (and enemies) will be for the rest of their lives…or at least for the next 285 days?
[Image: Flickr user Arman Thanvir]