Mining User Behavior in Social Recommender Systems
Carine Pierrette Mukamakuza, Dimitris Sacharidis, Hannes Werthner · 2018
Social recommender systems make use of the available information about social connections between users to improve the quality of the recommendations. The assumption is that if two users are connected, they are likely to have similar preferences, and thus the system should make similar recommendations. Recently many approaches have been proposed based around similar assumptions, whose validity however has not been systematically studied. In our work we make the first step towards examining whether there exist observable relationships between social connections and rating behavior in social recommenders. In particular, we examine publicly available datasets containing traces of rating behavior along with a social graph. Using techniques from social network analysis and statistics, we investigate whether heavy rates, having provided feedback on many items, are also popular, i.e., central in the social network, and vice versa. Our results indicate important connections between heaviness and popularity. Specifically, we find that heaviness implies popularity, and that the association is stronger among very heavy raters.