摘要
People consume items not only by themselves but also in groups that include their family, friends, coworkers, and online social groups. Different from individual recommendation systems, group recommendation systems first consider group member's preference separately. The impacts among members are considered, and then, the final group preference and decision would be generated. However, existing group recommendation models suffer severer data sparsity problems than traditional recommendation systems. There is currently lack of a systematic approach to properly address the above issue. What is worse, previous works less consider situations that users may have diverse interests, which means that users may change their preferences by considering the preferences of other group members. Here, we propose a member-augmented multi-interest model with knowledge graph (KG) embeddings to overcome the aforementioned drawbacks. Because only positive labels of groups can be identified in a dataset, precisely predicting a group's opinion on items that members have not been exposed to is difficult. Accordingly, our model applies the member augmentation (MA) technique to precisely predict a group's opinion on items. In addition, we leverage multi-interest framework to model the change of diverse user preferences. The framework can know which interests of the user will affect the decision of the group, and interests will vary with different group members. Experiments indicated that the proposed model improves the performance by around 20%, 10%, and 2% in MaFengWo, Yelp, and Meetup-NYC, respectively.