Research On Personalized Recommendation Algorithm Based On Knowledge Graph

In recent years,as an important method to solve the problem of information overload,the recommender system has been highly valued by researchers.However,data sparsity,cold start,accuracy and other issues still limit the performance of the recommender system.Therefore,researchers have proposed a solution to introduce side information into the recommendation system.As an important side information in the recommendation system,the knowledge graph has received extensive attention from researchers,which makes the personalized recommendation algorithm based on the knowledge graph has become a research hotspot nowadays.Therefore,the main research work of thesis is to propose two kinds of personalized recommendation algorithms based on knowledge graph for the scenarios of using explicit feedback data for scoring prediction and implicit feedback data for TOP-N recommendation.In the scenario of using explicit feedback data to predict user ratings,to solve the data sparsity problem and improve the accuracy of recommendation,the knowledge graph is introduced into the neural collaborative filtering model,and a neural collaborative filtering model assisted by the knowledge graph is proposed.The model uses a neural collaborative filtering model fused with generalized matrix factorization to construct a recommendation module,uses a deep semantic matching model to construct a knowledge representation module,and integrates the recommendation module and the knowledge representation module through the cross-connect module,and introduces the knowledge graph as side information into the recommendation system.Through the alternate training of the knowledge representation module of the recommendation module,the knowledge representation module assists the training of the recommendation module,which effectively improves the effect of rating prediction.Experiments show that this model not only improves the MAE and RMSE by 9.46% and 10.18% respectively than the User CF method,but also effectively alleviates the problem of data sparsity.In the scenario of using implicit feedback data for TOP-N recommendation,in order to solve the problem of the lack of negative feedback items of implicit feedback data and improve the accuracy of recommendation,a joint Bayesian ranking model of knowledge aware sampling is proposed.By improving the Bayesian Personalized Ranking algorithm,the model uses user item interaction data and inter-item joint purchase data to model the user's partial ordering relationship and functional complementary relationship,which improves the accuracy of recommendation.In addition,the model constructs a sampling strategy for knowledge aware,this strategy solves the problem of the lack of negative feedback items in implicit feedback data by considering the information and authenticity of negative feedback items.Experiments show that this model is 44.79%,43.80%,and38.93% higher than the Bayesian personalized sorting algorithm on Precision@10,Recall@10,and NDCG@10,respectively.