Research On Personalized Recommendation Algorithm Based On Knowledge Graph

Recommendation system is an important technology to solve information overload.However,the traditional recommendation model has the problems of data sparsity and cold start,which leads to the low recommendation accuracy.Research shows that the general solution is to improve the accuracy of recommendations by adding some additional semantic information.The knowledge graph contains a lot of semantic content,which can be used as auxiliary information of recommendation system.Graph convolutional neural network can deal with non-Euclidean structure,namely graph structure data.Combining knowledge graph with graph convolution neural network,graph convolution neural network is used to mine the features between nodes in the knowledge graph.In addition,user information can not be ignored in the recommendation system.Therefore,this thesis focuses on the personalized recommendation algorithm based on user information enhanced knowledge graph convolution.The research shows that the user's preference information can be extracted and added into the recommendation system to improve the accuracy of recommendation by using the historical interaction data between users and items.The main contents of this thesis are as follows:1.The traditional recommendation model has problems of data sparsity and low recommendation accuracy.This thesis combines knowledge graph with graph convolution neural network and proposes graph convolution network recommendation method(KMCN)based on knowledge graph.Graph convolutional neural network can automatically capture semantic information of higher-order structure in knowledge graph and compute node representation in graph.Firstly,the core idea of the algorithm is to use the edge structure information in the knowledge graph to carry out message propagation and neighborhood aggregation operation on nodes,so as to generate new node representation.Then,the interaction probability is obtained by inner product operation with the user vector.In neighborhood aggregation,a neighborhood node aggregation method is designed in this thesis.Firstly,the nodes with the same aggregation relationship are averaged,and then the nodes with different aggregation relationship are aggregated to obtain the final node vector representation through calculation.Finally,experiments are carried out on the well-known Movielens-1M and Book-Crossing datasets in the neighborhood of recommendation system,and the results show that this algorithm improves the accuracy of recommendation.2.In the recommendation process,user information should not be ignored.Because the KMCN algorithm does not fully consider the user information,but only extracts the features of items in the knowledge graph,the results of recommendation are relatively one-sided.It is found that the historical interaction information between users and items can be well represented as user preference information.Since the user vector and the corresponding entity vector in the knowledge graph are two different descriptions of the same object,they can obtain additional information needed by each other through cross-sharing.By introducing user information and extracting information between similar users,problems such as data sparsity and user cold start can be solved to a certain extent.Therefore,a personalized recommendation algorithm based on user information enhanced knowledge graph convolution(KMCNU)was proposed to improve the accuracy of recommendation.Firstly,the model is composed of four modules: information aggregation module,user information extraction module,cross link module and knowledge graph embedding module.Based on knowledge graph convolution recommendation,the algorithm introduces user information to improve the recommendation performance.Finally,compared with other models on Dianping-Food and Last.FM public data sets,the results show that the AUC of KMCNU model increases by 2.5% and 3.0%,respectively.