Knowledge Base Embeddings for a Recommendation Based on Overlapping Knowledge and Graph Learning

Knowledge bases have grown in size and complexity due to the rapid development of information technology, and recommendation algorithms are the key to solving the information overload problem caused by massive knowledge. Generating knowledge base embeddings by discovering potential associations between users and knowledge can improve recommendation effectiveness, but current approaches fail to fully exploit the overlapping knowledge among users. To enhance knowledge recommendations’ effectiveness, this paper presents a novel Knowledge Base embedding technique based on overlapping knowledge and graph learning. The proposed method is divided into a Graph learning module and a Semantic injection module. The Graph learning module utilizes the overlapping knowledge between different users. It uses a novel Weighted Hyper Sub Graph with a twin-layer structure to generate the KB embedding by a modified Graph Neural Network (GNN) model. The Semantic injection module automatically extracts semantic information through a pre-trained language model based on BERT. The final recommendation result was generated by joining the Graph learning and Semantic injection modules together. The results of our experiments on Citeseer, DBLP and CiteUlik-a datasets demonstrate that our method improves performance on the knowledge recommendation task and successfully captures overlapping user knowledge.