Group Recommendation Algorithms Based On Graph Neural Networks

As group activities become more and more common on the Internet,how to recommend an item or an activity to a group has attracted the attention of many scholars.Due to the complex relationship within the group,personalized recommendation algorithms unable to recommend items that satisfy the entire group.Therefore,we need more specific group recommendation algorithms.Most of the existing group recommendation work obtain the group's preferences by integrating the preferences of the members in the group.These approaches ignore the behaviorial similarity of groups.Recently,the graph neural network has shown amazing potential in personalized recommendation systems.The graph neural network encodes the similarity of nodes through the embedding propagation process,which helps to obtain effective users' and items' representations.This paper applied graph neural network to group recommendation tasks.The main research work of this paper are as follows:(1)This paper applied the graph convolution network to the group-user-item graph,and proposed a group recommendation framework based on a single graph convolution network.Experiments are conducted on multiple datasets to verify the effectiveness of the framework and the impact of different embedding propagation methods on group recommendation.(2)Considering that the group-user-item graph contains different types of edges,it is easy to generate noise when the different types of edges are merged together,so this paper divided the group-user-item graph into three subgraphs,then performed embedding propagation on the subgraphs respectively.Experiments on the same datasets verified that the splitting of the graph can improve the group recommendation quality.(3)From another perspective,considering the heterogeneity of the group-user-item graph,This paper introduced the heterogeneous graph attention network into the group recommendation task,and proposed a hierarchical attention structure to distinguish the impact of different nodes and different edges,respectively.The experimental results proved that the heterogeneous graph attention network can achieve accurate group recommendation.