Research On Graph Neural Network Recommendation Algorithm Based On Knowledge Graph

Recommender systems play a key role in alleviating information overload,and classical collaborative filtering recommendation algorithm uses the users’ historical interaction records to mine user preferences.However,collaborative filtering recommendation algorithm usually suffer from sparsity and cold start problems.Knowledge graph is introduced as side information into recommender systems,which can alleviate the above problems and has been widely concerned.Graph neural network has excellent learning ability for graph structure data and have become an important method in the field of recommendation.Therefore,graph neural network recommendation algorithm based on knowledge graph have been a hot research topic in recent years.However,existing studies have ignored the adverse effects of noise on recommendation performance.The redundant noise information of the knowledge graph and user noise interactions bring challenges to recommendation.In addition,user history interaction data is sparse,existing studies only focus on explicit modeling of user-item interactions,and graph neural network can only aggregate local neighborhood information of items,making it difficult to extract non-local neighborhood information,which leads to insufficient mining of users’ potential interests and limits the recommendation performance.Therefore,this thesis conducts research on graph neural network recommendation algorithm based on knowledge graph,and proposes dual-channel graph neural networks based on knowledge graph for denoising recommendation and graph neural networks based on potential knowledge enhancement for recommendation algorithm.The main research contents are as follows:(1)To address the noise problem in the recommendation process,this thesis proposes dual-channel graph neural networks based on knowledge graph(KDGNN)to improve the recommendation performance by reducing the noise in the recommendation process.Firstly,a dual-channel graph neural networks is designed,which can extract both neighbor entity and relational features,so as to extract knowledge information from knowledge graph to the maximum extent.Secondly,the algorithm designs a personalized gating mechanism,namely dual-channel balancing mechanism,to reduce the propagation of redundant noise in the knowledge graph.Then,user personalized and knowledge-aware signals are integrated to capture user preferences fully,use personalized knowledge-aware attention to denoise user-item interaction data.Finally,experimental performance analysis is performed on three public datasets to verify the effectiveness of the KDGNN algorithm.The experimental results show that the KDGNN algorithm outperforms the compared advanced baseline algorithms,especially on the Last.FM dataset,with an improvement of2.6% on AUC and 1.9% on F1.(2)To address the problem that the user interaction data is sparse and the existing methods are difficult to mine the non-local knowledge information of the item.This thesis proposes a graph neural network recommendation algorithm based on potential knowledge enhancement(PKGNN).Specifically,firstly,the user-item interaction data is fully utilized to extend the users’ potential interest items.Then,the rich semantic information in the knowledge graph is used to extend the potential items that have semantic relevance to the items.The potentially related items of users and items are propagated on the knowledge graph to extract the non-local neighborhood information of users and items.Finally,this thesis conducts experimental performance analysis on different datasets.The experimental results show that PKGNN has better recommendation performance compared with the baseline algorithm,and the overall performance is improved by about 5% compared with the strongest baseline algorithm.