Multi-relation Neural Network Recommendation Model Based on Knowledge Graph Embedding Algorithm

Most of traditional collaborative filtering recommendation system (RS) suffers from cold start problem. In recent years, knowledge graph (KG) has become a useful tool to extract features from multi-relation graphs. Therefore, more and more researches begin to fuse KG into RS, which is also called KG-based RS. Unfortunately, most of existing KG based methods fail to make full use of knowledge graph to extract potential relations between entities. To address these shortcomings, we propose an enhanced multi-task recommendation framework (TransD-based RippleNet). In this framework, we introduce knowledge graph embedding (KGE) algorithm TransD into RS, which can extract the heterogeneity of implicit information. The user-item knowledge graph was pre-trained in TransD model in the beginning and then the results will be sent into an improved RippleNet, reducing the computational time. Extensive experiments on muti-task recommendation datasets are conducted on the proposed framework. The results show that the proposed framework can converge quicker, while attaining a better recommendation accuracy in multi-task scenario.