Enhancing signed social recommendation via extracting consistent and inconsistent relations

Signed Social recommendations leverage signed social information(e.g., trust and distrust) to alleviate the cold-start and data sparsity problem. Recently, Graph Neural Network (GNN) methods have demonstrated the powerful in graph representation learning, which motivates GNN-based social recommendation frameworks. However, building GNN-based signed social recommender systems faces challenges. For example, signed social recommendations face the social inconsistency problem, which indicates that the evidence of item preferences provided by the social information and user-item interactions information are not necessarily consistent. In order to alleviate social inconsistency problem, we present a novel GNN-based E nhancing S igned S ocial Rec ommendation framework(ESSRec). Specifically, ESSRec first learns item-space user embedding and final item embedding by embedding propagation on the user-item graph. Then, it reconstructs the signed social graph by extracting consistently positive relations, consistently negative relations, and inconsistent relations from original graph based on the item-space user embedding. Moreover, we design the embedding propagation rule on the reconstructed signed social graph to empowered GNNs Model. Extensive experiments on real-world dataset Epinions demonstrate the effectiveness of the proposed framework ESSRec, with more than 8% on MAE and 4% on RMSE performance improvements over the best baseline for rating prediction. Further experiments demonstrate that extracting consistent and inconsistent relations and reconstructing signed social networks can improve the performance of ESSRec.