Research On Graph Neural Network-based Knowledge Graph Recommendation Algorithm

Traditional recommendation algorithms are mainly based on user-item interaction data and use collaborative filtering to make recommendations.When faced with sparse user-item interaction data,the recommendation performance is greatly affected by the lack of collaborative signals.An effective solution is to design a suitable graph neural network framework to introduce knowledge graph information,but there are still shortcomings in information modeling and fusion.Therefore,in this paper,based on user-item interaction data and knowledge graph construction graph structure data,we propose an improved knowledge graph recommendation algorithm using graph neural network technology in terms of relationship modeling and fusion of knowledge information and collaborative information in the knowledge graph.Firstly,in terms of knowledge graph relationship modeling,this paper discusses the shortcomings brought by existing research modeling relationships as scalars and proposes a knowledge graph recommendation algorithm for vectorizing relationships.The algorithm is based on the graph neural network framework,which represents the relationships in the knowledge graph quantitatively and improves the information propagation mechanism so that the high-order knowledge information in the knowledge graph can be learned more accurately.In addition,in the negative sampling stage of the algorithm,this paper improves the random negative sampling into relationship-oriented negative sampling to further optimize the learning process of the algorithm.Secondly,in the fusion of collaborative information and knowledge information,this paper discusses the problems caused by the existing research focusing on knowledge information and the insufficient fusion with collaborative information,and proposes a recommendation algorithm that fuses knowledge and collaborative information.Based on the graph neural network framework,the algorithm designs a two-channel information propagation aggregation mechanism to generate user and item representations containing knowledge information and collaborative respectively,and weighted fusion of the two by the attention mechanism.Finally,this paper conducts experiments on three publicly available datasets,Amazon-book,Last-FM,Alibaba-i Fashion,and compares them with baseline methods such as KGNN-LS,KGAT,CKE,etc.The key modules and important parameters are explored and studied.The effectiveness of the algorithm in modeling knowledge graph information to improve recommendation effectiveness is analyzed and verified.This paper provides new ideas in the introduction of knowledge graph information to enhance the accuracy of recommendations,which has theoretical and practical significance.