Integrating social networks in a recommender system for movies

This thesis project was done for Ericsson Research in Stockholm, Sweden. The purpose was to evaluate how well an existing algorithm in a recommender system predicts movie ratings and get an indication of how the users perceive the recommendations given by the system. The recommendations are computed with a revised User-based Collaborative Filtering algorithm that calculates trust amongst people in a social network to identify the most suited recommenders. The purpose has also been to use friends from a social networking site as the social network and one of the project goals was to build an application on such a site. The prototyping activities resulted in a Facebook application for personalized movie recommendations called MovieBuddy, where a user could rate movies and choose friends that would influence their recommendations. The recommender system was evaluated from three main aspects: the accuracy of predicted ratings compared to users actual ratings, the accuracy of inferred trust values compared to users actual trust ratings and the users' opinions as gathered from a survey regarding their perception of the recommendations. The results showed that the users rated the movies in the recommendation list higher than they did anywhere else. Still, both the inferred trust values and the predicted ratings had quite large mean absolute errors and while the users were overall positive to the recommender system, they felt that there was room for improvement when it came to the recommendations. According to the survey, users were interested in more user tasks than the application supported, like being able to filter and search for movies. Overall, they were less interested in rating movies and more interested in what the system could do for them. The results lead to some conclusions on how the recommendation algorithm could be revised and how the system could use social networking sites as a source of information and inspiration.