Implicit-Relation Knowledge Distillation for Robust Recommendations

Wenhao Liu, Sui Lin, Lanjun Li

Abstract: Social recommendation systems utilize social connections to build recommender systems aimed at alleviating information overload. However, existing methods often fail to fully capture the rich information embedded in heterogeneous relationships, which are crucial for enhancing user preference learning. To tackle these challenges, we propose the IDRec method, which employs an Implicit-relation knowledge Distillation approach to integrate information from both the interaction graph and the social graph. This involves training two auxiliary models, each utilizing one of the graphs, and enhancing the heterogeneous graph model with meta networks for personalized knowledge transformation and adaptive augmentation. Real-life recommendation scenarios involve heterogeneous relationships that provide valuable information for enhancing user preference learning. By incorporating these diverse relational semantics, IDRec effectively captures the intricate dependencies within the data. Extensive experiments conducted on three open datasets demonstrate the superiority of IDRec over state-of-the-art methods. The results show that IDRec significantly improves recommender system performance by integrating comprehensive social and interaction data and leveraging meta networks for a more personalized and adaptive recommendation process. These experimental results validate the effectiveness of IDRec in enhancing recommendation accuracy and robustness compared to existing methods.