Improving Recommendations from Knowledge Graphs by Multi-Level Semantic Enrichment Based on Clustering and Attention Mechanism

Abstract

The increasing popularity of recommender systems is leading to a focus on improving them using knowledge graphs (KG). This new trend has emerged to address the problem of data sparsity. It is evident that hybrid approaches combining collaborative and content-based filtering have demonstrated considerable potential; however, they do not fully use the semantic richness of knowledge graphs. To address this, we propose a method that enriches the knowledge graph and improves the recommendation performance by identifying implicit relationships between entities using unsupervised clustering and cosine similarity calculations. These relationships are then integrated into the graph in the form of new weighted edges. The TransR model is then used to recalculate the enriched embeddings, before employing an attention mechanism to capture complex dependencies between users and items. Experiments conducted on the Yelp2018 and Amazon datasets demonstrate a substantial enhancement in the efficacy of recommendation systems, underscoring the efficacy of the proposed approach in enhancing attentional propagation and generating recommendations that are more pertinent and tailored to individual users.