Knowledge graph fine-grained network with attribute transfer for recommendation

Abstract

In recent times, knowledge graph (KG) has been frequently incorporated into recommendation systems as side information. The end-to-end method based on graph neural network (GNN), representing a contemporary technical hotspot, can comprehensively mine KG information. However, most GNN-based models (1) primarily emphasize the relative signals between users and items, and the collaborative signals among the users, but overlook the collaborative signals among the items; (2) generally struggle to effectively distinguish between cold-start items and general items in embedding learning. These drawbacks will cause insufficient information and low accuracy of item representations, further impairing the interpretability and accuracy of recommendation. To tackle them, Knowledge graph fine-grained network with attribute transfer for recommendation (KGFA) is proposed in this paper to address the insufficient correlations mining among items and inadequate cold-start item representation information in the knowledge-aware recommendation. Specifically, our method adopts attentive combination of intents to model user representation at a fine-grained level, and differentiates the contribution of diverse intents to user behavior for better model interpretability. Moreover, the cold-start items’ neighbors are further propagated in our meticulously designed item attribute transfer space to thoroughly excavate the collaborative signals between items and enrich the information in the cold-start item representations. Abundant experiments on two public datasets verify the outstanding effectiveness of KGFA over state-of-the-art methods.