Group-based social diffusion in recommendation

In social-enhanced recommendation systems such as Twitter and Weibo, users could get information from both personalized recommendation and social diffusion modules. In real-world scenarios, the user-group-user based social diffusion plays an essential role to efficiently broadcast information to groups of target users. Through this diffusion path, users first click items provided by the recommendation module, and then share the clicked items to the target user groups. Other users in the group can click the shared items, and return back to the recommendation module for more contents and related items. However, most social-enhanced recommendation systems merely focus on the recommendation module that they can directly influence, ignoring explicitly modeling and predicting for the social diffusion module. In this work, we propose a novel Group-based social diffusion (GSD) model, which aims to jointly optimize the click, share, and return stages in social-enhanced recommendation. We design a heterogeneous ternary graph neural network to jointly model the complex binary and ternary relations among users, items, and groups. We conduct extensive experiments and achieve significant improvements on all click, share, and return prediction tasks, and also achieve promising results on a new full-chain social impact prediction task.