Semantics-guided disentangled learning for recommendation

Although traditional recommendation methods trained on observational interaction information have engendered a significant impact in real-world applications, it is challenging to disentangle users’ true interests from interaction data. Recent disentangled learning methods emphasize on untangling users’ true interests from historical interaction records, which however overlook auxiliary information to correct bias. In this paper, we design a novel method called SeDLR (Semantics Disentangled Learning Recommendation) to bridge this gap. Particularly, by leveraging rich heterogeneous information networks (HIN), SeDLR is capable of untangling high-order user-item relationships into multiple independent components according to their semantic user intents. In addition, SeDLR offers reliable explanations for the disentangled graph embeddings by the designed Monte Carlo edge-drop component. Finally, we conduct extensive experiments on two benchmark datasets and achieve state-of-the-art performance compared against recent strong baselines. Copyright © 2022 The Author(s), under exclusive license to Springer Nature Switzerland AG.