Quad-tier entity fusion contrastive representation learning for knowledge aware recommendation system

Abstract

Knowledge graph (KG) has recently emerged as a powerful source of auxiliary information in the realm of knowledge-aware recommendation (KGR) systems. However, due to the lack of supervision signals caused by the sparse nature of user-item interactions, existing supervised graph neural network (GNN) models suffer from performance degradation. Moreover, the over-smoothing issue further limits the number of GNN layers or hops required to propagate messages-these models ignore the non-local information concealed deep within the knowledge graph. We propose the Quad-Tier Entity Fusion Contrastive Representation Learning (QTEF-CRL) knowledge-aware framework to achieve learning of deep user preferences from four perspectives: the collaborative, semantic, preference, and structural view. Unlike existing methods, the proposed tri-local and single-global quad-tier architecture exploits the knowledge graph holistically to achieve effective self-supervised representation learning. The newly-introduced preference view constructed from the collaborative knowledge graph (CKG) comprises a preference graph and preference-guided GNN that are specifically designed to capture non-local information explicitly. Experiments conducted on three datasets highlight the efficacy of our proposed model.