RFAN: Relation-fused multi-head attention network for knowledge graph enhanced recommendation

Knowledge graphs (KGs) provide rich external structures and semantic information for recommendation systems and have attracted extensive attention recently. Some existing KG-based recommendation methods process triplets independently through knowledge graph embedding, and they fail to fully capture the complex and implicit relation information in the neighbourhood around a given entity. To solve this problem, we borrow Graph Attention Network (GAT) and propose a relation-fused multi-head attention network (RFAN) that integrates relations into an attention network in a KG to learn the representation of entities through different contributions of neighbours and then combine them with user-item interactions to capture user preferences. In addition, we use a non-sampling strategy to learn parameters from the whole training data to improve training efficiency. We apply the proposed model on three public benchmarks, and the experimental results show that the RFAN outperforms several state-of-the-art methods.

Automated summary

Knowledge graphs (KGs) provide rich external structures and semantic information for recommendation systems. Existing methods process triplets independently, failing to capture the complex and implicit relation information. To address this, we propose a relation-fused multi-head attention network (RFAN) that integrates relations into an attention network in a KG, effectively capturing user preferences. Experimental results demonstrate that RFAN outperforms other methods.