Research On Knowledge Graph Recommendation Algorithm Based On Graph Neural Network

Knowledge graph has been introduced into the personalized recommendation algorithm for learning users’ preferences due to its strong ability of structure information expression and edge information mining.Graph neural network technology is an important method in the field of knowledge graph recommendation.However,the graph neural network has the over-smoothing problem,which makes it difficult to construct a deeper network to learn users’ long-range interests and brings challenges to recommendation.In addition,most entities in the knowledge graph have merely a small number of triples,resulting in a long-tail phenomenon,which leads to a small amount of data available for training,which limits the performance of the recommender system.Therefore,the paper proposes two kinds of knowledge graph recommendation algorithm based on the graph neural network to solve the over-smoothing problem of graph neural network and the long tail problem of knowledge graph,respectively.The details are as follows:(1)The knowledge graph random neural network for recommender systems is proposed to solve over-smoothing problem.A random feature propagation is designed to learn feature representations of items.First,the disturbed entity matrix is generated by random dropout strategy.Then,feature propagation is carried out based on disturbed feature matrix to obtain higher-order neighborhood information,locating the novel entity embedding representation,and generate augmented feature matrix.Finally,the consistency regularization method is used to optimize the prediction results on multiple augmented feature matrices generated after executing the random dropout strategy for several times.(2)The knowledge graph extrapolation network with relation learning for recommendation is proposed to solve long-tail problem.Firstly,the embedding representations of users and items are learned by the knowledge propagation combining collaborative signal,and attention mechanism is designed to distinguish the contributions of different neighbor nodes in the propagation.Then,the relation learning is designed to learn relation between unknown items.The learned embedding of items is input through two individual knowledge propagation layers to obtain the novel representations.The relationship between unknown items is modeled on the novel representation of items to enrich the feature representations of items.In this paper,a large number of experiments are conducted on real recommendation datasets.The experimental results show that the recommendation algorithm based on knowledge graph random neural network effectively alleviate the over-smoothing problem and improve the accuracy of user preference prediction.The knowledge graph extrapolation network with relation learning for recommendation achieves accurate and stable recommendation in the scenario of long tail phenomenon.Therefore,the two algorithms are proposed in this paper have ability to accomplish personalized recommendation task effectively.