Similarity attributed knowledge graph embedding enhancement for item recommendation

Abstract

Knowledge Graph Embedding (KGE)-enhanced recommender systems are effective in providing accurate and personalized recommendations in diverse application scenarios. However, such techniques that exploit entire embedded Knowledge Graph (KG) without data relevance approval constraints fail to stop noise penetration into the data. Additionally, approaches that pay no heed to tackle semantic relations among entities remain unable to effectively capture semantical structure of Heterogeneous Information Graph (HIG). Therefore, in this paper, we propose Similarity Attributed Graph-embedding Enhancement (SAGE) approach to model similarity-aware semantic connections among entities according to their triplets’ granularity. SAGE is a novel Knowledge Graph Embedding Enhancement (KGEE) method that constructs Entity-relevance-based Similarity-attributed Subgraph (ESS) to remove noise from the underlying data. It propagates interactions-enhanced knowledge over ESS to learn higher-order semantic connections among entities; and simultaneously utilizes feedbacks to enhance the interactions and regularize the model to highlight influential targets (nodes). Further, it samples influential targets in KG, independently move their preferences to the Local Central Nodes (LCN) of current influential areas, and streamline the collected information from all LCN to the main unit. Finally, a prediction module is used to determine generalized preferences for recommendation. We performed extensive experiments on benchmark datasets to evaluate the performance of SAGE where it outperformed the state-of-the-art methods with significant improvements in effectively providing the desired explainable recommendations.