It is often assumed that our online social interactions are close to random. New research from HP Labs, however, suggests that they’re actually predictable to a surprising degree.
By using techniques that were first developed to predict genetic sequences, HP researchers have shown that our past social networking activities can be used to make accurate predictions about our future interactions – with obvious consequences for improving the design of social networks.
The research, conducted by HP Labs’ Bernardo Huberman and Stanford University PhD student Chunyan Wang and available here, examined datasets from the websites Epinions and Whrrl featuring hundreds of thousands of user comments and check-ins to explore the social behavior of people acting both as individuals and as members of online groups.
“Interestingly, we found that individuals are slightly less predictable when they are acting as part of a group than when they are acting alone,” says Huberman, Senior HP Fellow and director of the Social Computing Group at HP Labs. The finding, he notes, contrasts with the assumptions behind many models of online group behavior, which suggest that group behavior ought to be easier to predict.
In the past, adds Huberman, we’ve made predictions about what people are really searching for, or about what they’d be interested in buying, by comparing their interests with those of millions of other people who are more or less like them.
That statistical approach has proven useful for search, online recommendations, link prediction and advertising. However, he says, “it’s not clear that this approach works well as a predictor in more interactive processes such as contacting friends within social networks, participating in online discourse and exchanges of email and text messages.”
As Huberman and Wang’s paper shows, it turns out that one can predict an individual’s future social actions based solely on that person’s past social behavior.
“It’s not that we can now perfectly predict what people are going to do,” explains Huberman. “But we are showing that our social interactions online aren’t at all random and we’re suggesting that it would be wrong to assume that they are.”
Incorporating this new understanding ought to help us better analyze the vast sets of social data now available to us, and to better predict how people interact with each other. That, in turn, should lead to improved social media and services. It could even be used to improve the statistical algorithms that currently dominate recommendations, advertising and search.
“I can see HP using this to tap into the way people talk about our products,” Huberman suggests. “And it applies even to interactions inside the enterprise. We could predict fairly accurately how employees will interact with each other, for example, and what kind of resources they will make use of through the internet. That’s can to have a positive impact on an enterprise’s ability to run its own processes and to reach customers more efficiently.”