Share on Facebook Tweet on Twitter Share on LinkedIn Share by email
Prediction Square

This project is exploring a new dimension for social search based on eliciting personalised predictions about a user's preferences from social contacts.


You can try our first experimental version of Prediction Square on Facebook at

Today’s social search and recommendation systems are generally based on the friends’ opinion and usage patterns paradigms, which try to predict people’s taste using the assumptions that “your friends are like you” or that “people who agree on some things also agree on others”. There is of course a limit to the effectiveness of such techniques in practice. Firstly, people in the social network of the user are usually family, colleagues, or casual acquaintances, and not necessarily people with tastes, opinions or circumstances similar to the user. Even if friends share certain opinions, there may be many things on which friends disagree given their individual personalities. The same is true when analysing usage patterns: agreeing on certain items does not always mean agreeing on others. People can be very similar in many ways, but also very unique in many others, and this uniqueness is what is not addressed by current methods that classify users with their friends or other users who had matching opinions in the past.

In this project we are experimenting with a new signal for social/personalised search which is aimed at addressing the individuality of people, based on the idea that “your friends know you” rather than "your friends are like you". The method involves eliciting personalised predictions about a user’s taste from friends, who may know the user well, even if they do not always share the same opinions. The incentive for making predictions is to receive scores and feedback as to how accurate one's predictions are with respect to actual opinions, and one's "prediction ability" in general. We are currently testing this approach with a simple application deployed on the Facebook platform, which is aimed at collecting data and user feedback.