GRUIFI: A Group Recommendation Model Covering User Importance and Feature Interaction

Abstract

Group recommendation derives from a phenomenon that a group with similar interests have formed various communities, which creates the requirements that a group of users in one community want to share personalized services. Different from traditional recommendations that focus on individuals, group recommendation needs to consider the differences in preference of group members. How to build a proper model for group members to aggregate different preferences is still a challenging problem: (1) the influence of group members is quite different; (2) a user decision is directly or indirectly influenced by other members in the same group. This paper proposed a Group Recommendation model covering User Importance and automatic Feature Interaction (GRUIFI), which can model interaction data of group member and learn group potential preference representation. Our model exploits an attention mechanism to obtain the weights of group members that represent user importance, and those dynamic user weights are integrated to learn a group representation. Then we design a neural network that combines the multi-head attention to automatically learn fine-grained interactions between groups and items, and further capture the interdependency between group members. Finally, the experiments on the two real-world datasets show that GRUIFI performs significantly better than baseline methods.