TrustGNN: Enhancing GNN via Multi-similarity Neighbors Identifying for Social Recommendation
Qi Han · 2022 IEEE Conference on Telecommunications, Optics and Computer Science (TOCS) · 2022
Recently, with the development of Graph Neural Networks (GNN), many GNN-based methods have been developed for recommendation systems to aggregate both user-user and user-item interactions. However, most existing GNN-based models only consider direct relationship during the social aggregation process of user modeling, making the users selected as ‘neighbors’ can’t well reflect target user’s preferences and latent factors. To take more elements (such as the propagation of trust) into account during neighbors recognition and obtain a better user latent factor, we present a framework (TrustGNN), which employs DeepWalk method to help obtain the implicit similarity in one’s trustees, and combines it with the explicit similarity of a certain user’s historical ratings. We identify user’s neighbors according to the combined similarity, and then apply them in the social aggregation process. The method models two graphs and considers heterogeneous strengths. We test our proposed model on two real-life datasets, Filmtrust and CiaoDVD. The experiments demonstrate the effectiveness of our proposed model.