Research On User Perference Recommendation Model Integrating Knowledge Graph

With the wide application of Internet and the explosive growth of information,the recommendation system has become an important technology in our daily lives.It can not only filter out some useless information,but also realize personalized recommendations for different users.Traditional recommendation systems usually recommend through the interaction between users and items,such as the user's rating for items.However,these simple rating numbers are difficult to reflect the subjectivity of user preferences and the interpretability of user choices,and sparse data will also affect the accuracy of recommendation results.In order to solve the above problems,this paper proposes a user preference recommendation model which fused with knowledge graphs.The main works are as follows:First of all,this paper proposes a recommendation method combining tag and knowledge graph for recommendation(CTK).In this method,the semantic information contained in the tags marked by users is used to mine users' preferences and construct user features.It also uses a large amount of entity information in the knowledge graph and uses the KGCNN model(Knowledge graph Grouping Convolutional Neural Network,KGCNN)to update fine-grained characterization of item features.Finally,jointly calculate user features and item features thus improving the accuracy,personalization and interpretability of recommendations.Secondly,this paper proposes a hybrid attention model based on CTK model(CTK-HATT),which uses the attention mechanism to calculate the user's preference for items hidden in the user's tag,and uses the self-attention mechanism to mine the different emphasis of the item's own characteristics.In combination of the two,hybrid attention weights are assigned to different features,and more accurate and comprehensive item features are constructed.Finally,according to the time sequence of the interaction between users and items,this paper proposes a recommendation method LSPR(Combing Long-term and Shortterm Preference of user for Recommendation)to explore the changes of users' preference over time.LSPR model first uses the CTK-HATT model to mine the user's long-term preferences;then uses user preference prediction and Hidden Markov model to mine users' short-term preferences;Finally,the two are integrated into make recommendations for users,which not only contains the overall characteristics of users,but also includes the time series characteristics of user behavior,so as to improve the effects of the recommendation.In experimental part,we conduct experiment on Movie Lens 10 M and use MAE,MSE and AUC as metrics.Compared with other traditional recommendation models,all metrics of our proposed model have been improved.