Matrix Factorization in Social Group Recommender Systems

Abstract

Traditionally, Group Recommender Systems (GRS) apply an aggregation approach, which computes a group rating for each item by estimating unknown individual ratings, for which has been demonstrated that matrix factorization (MF) models are superior to classic nearest-neighbor techniques in individual recommender systems. Moreover, when people are in a group making a choice from alternatives, they tend to change their opinions accordingly to the social influence exerted by others' group members. Sociological analyses suggest that some social factors express social influence in a group, such as, cohesion, social similarity and social centrality. In this work, we combine a MF model to estimate unknown ratings with a social network analysis (SNA) to evidence possible social influence. Firstly, we present an analysis of the relevance of social factors detected in relation with the members' opinions and, then, we describe the results obtained when comparing the proposed technique with the classic group recommender technique.