Preference relation based collaborative filtering with graph aggregation for group recommender system

Abstract

Most of the group recommender systems (GRS) apply some aggregation strategy to the ratings given by the group members for generating recommendations. But this can be highly influenced by a few members of the group, which can lead to poor group recommendation. Further, rating based aggregation strategies do not provide efficient ranking of items. Keeping these things in mind, this paper proposes a preference relation (PR) based GRS, that uses matrix factorization (MF) for predicting unknown PRs for group members. The aggregation of preferences is done using a novel virtual user based weight aggregation strategy. The weight aggregation concept is derived from the graph aggregation process. The advantage of this process is that it does not ignore weak preferences and also contributes towards group recommendation. The proposed model is evaluated and compared using standard ranking measures for MovieLens and NetFlix datasets. Experimental results obtained using Top-K recommendation task indicates the superiority of the proposed GRS method over the others. The proposed GRS model provides the best performance when we balance the number of member in a group and the number of recommended items.