Gadgets That Know Your Next Move
Researchers have developed a model that predicts people’s daily activities
Source: “Eigenbehaviors: Identifying Structure in Routine”
Nathan Eagle et al.
MIT Media Lab Vision and Modeling Technical Report 601
Results: Using location data, call logs, and other information collected from mobile phones, Nathan Eagle and Sandy Pentland of MIT’s Media Laboratory have developed a new data-analysis technique that, with only limited initial information, can predict the daily behavior and determine the social allegiances of study participants. By looking at a few early-morning activities and locations, the researchers can forecast a person’s remaining daily activities, associations, and locations with 79 percent accuracy. They can also identify group affiliations with 96 percent accuracy.
Why It Matters: As mobile devices generate increasingly immense amounts of behavioral data–about whom we call, where we go, and who is around us–they could learn to schedule meetings or recommend activities. But that will require new techniques to make sense of the data. Current computer models that predict behavior are complex and sometimes miss patterns that are simple for humans to see. The researchers’ approach can characterize and predict behavior more easily.
Methods: During the 2004-2005 school year, the researchers logged more than 350,000 hours of behavioral data collected from the mobile phones of 100 students and faculty members at MIT. The data included information on where participants were, whom they talked to on the phone, and which other participants were nearby. From this information, Eagle and Pentland extracted fundamental patterns–dubbed eigenbehaviors–that succinctly describe a person’s or a group’s daily activities. For example, sleeping late in the morning is part of the same eigenbehavior as going out that evening. Although the connection between these two behaviors may seem obvious to a person, it is difficult for a computer to spot using traditional behavior-prediction models.
Next Steps: The researchers are looking beyond individual behaviors and group affiliations to explore people’s influences on one another. They will test how well they can determine the satisfaction of people working on projects in groups, with an eye toward predicting which groups will be more efficient.