Facebook interactions utilization for addressing recommender systems cold start problem across system domain

Abstract

Recommender systems face cold start problem at the time of inception which researchers have addressed by transferring rating knowledge from auxiliary to target domain, hence laying the foundation of cross domain recommender systems (CDRS). Recently social media has been utilized as a potential auxiliary domain for recommender systems because they provide interactions such as likes, read, download, play, click etc., for related items. Among existing social media, Facebook provides rich social interactions, broadly grouped as private and public social interactions. Although research exists on private Facebook social interactions, Facebook public social interactions were not explored before. In this paper, we propose Facebook Direct Social Recommendation (Facebook DSR) approach which transforms Facebook's public interactions related to a social page into rating matrix and generate recommendations for provided list of items. Relative ratings obtained from proposed algorithm and random approach were compared with Movielens ratings. It was observed that proposed approach outperformed random approach threefold with respect to mean absolute error (MAE) evaluation and 19.4 % with respect to precision evaluation for new user cold start scenario. Finally, this study highlights future directions.