Knowledge Graph-enhanced Recommendation Methods Via Heterogeneous Network Representation Learning

With the development of artificial intelligence and the exponential growth of Internet data,people often confronted with information overload problem.It is difficult for users to quickly obtain the desired information from massive data,this makes intelligent and effective recommendation methods and recommendation systems more and more important.Among them,the recommendation method based on knowledge graph introduces the rich semantic relationships contained in users and items,which can effectively alleviate the data sparseness and cold start problems faced by the recommendation system,and greatly improve the effect of the recommendation model,and has been received by many scholars.Although the introduction of knowledge graph has greatly enhanced the content we can learn,the existing models are studying how to effectively integrate the knowledge graph into the user interaction map,without considering the graph pollution caused by the introduction of additional information.This pollution may cause the learned node representation to have neither recommendation task attribute nor knowledge graph representation learning task attribute.In addition,due to the indistinguishability of graph,the existing models do not distinguish between recommendation task and knowledge graph representation learning task,and learn the two task features at the same time through one method.Based on the above problems,the main work of this paper is as follows:(1)A new recommendation method based on knowledge graph fusion is proposed to deal with the image pollution caused by the inconsistency of information.Specifically,according to the characteristics of knowledge graph and user project interaction graph,this paper constructs a heterogeneous collaborative knowledge graph.The knowledge graph retains the original heterogeneous network side information of the knowledge graph,and adds new user interaction type information.The project is connected to the knowledge graph through entity matching link,and finally forms a heterogeneous collaborative knowledge graph.In addition,based on the heterogeneous collaborative knowledge graph,the heterogeneous network representation learning method can be easily used to complete the recommendation task,avoiding the computational burden caused by learning entity semantics on a large-scale knowledge graph.(2)On the basis of heterogeneous collaborative knowledge graph,this study designs a heterogeneous graph attention network which can distinguish recommendation task and knowledge graph representation learning task.By using different graph neural network learning methods on heterogeneous graph,not only the entity relationship of knowledge graph is preserved,but also the particularity of recommendation task is considered.Moreover,in the part of recommendation task,this study designs a new attention propagation method which can consider the interaction between users and projects,and encodes the interaction information of users and projects into the node representation.Finally,the effectiveness of the proposed method is verified on Amazon-book dataset.