Preference Networks And Non-Linear Preferences In Group Recommendations

Group recommender systems generate recommendations for a group by aggregating individual members' preferences and finding items that are liked by most of the members. In this paper we introduce a new approach to preference aggregation and group choice prediction that is based on a new form of weighting individuals' preferences. The approach is based on network science, and, in particular, it relies on the computation of node centrality scores in preferences similarity networks of groups. We also motivate and introduce a non-linear (exponential) remapping of the individuals' preferences. Based on offline experiments we demonstrate: 1) non-linear remapping of preferences is useful to better predict group choices and generate recommendations; and 2) our weighted approach predicts the actual group choices more accurately than current state-of-the-art methods for group recommendations.