Research On Personalized Recommendation Method Based On Knowledge Graph And Deep Learning

With the advent of the information era,the total amount of information has greatly exceeded the scope of personal acceptance,processing or effective use.In view of the endless and complicated information,how to distinguish the value of information and how to get the information that users really need is a question worthy of consideration and extremely urgent.The emergence of recommender systems effectively solves the problem of information overload.It can filter the useless information for users in an intelligent and automatic way and meet the needs of users at the same time,thus reducing the time and energy for users to obtain information.Although the traditional recommendation methods,such as collaborative filtering and content-based,have achieved good performance to some extent,the performance of the recommender systems will be significantly reduced in the face of new users or items due to the lack of feature information of users or items.In recent years,more and more researches try to introduce knowledge graph as auxiliary information into the recommender systems,but the existing knowledge graph aware recommendation methods still do not make full use of the feature and structural information of knowledge graph and cannot combine the knowledge graph representation learning with recommendation task effectively.To address the above limitations,this paper proposes UPPM and UIKJR models based on knowledge graph and deep learning technology.This paper is mainly as follows:(1)In this paper,an end-to-end user personalized preference modeling method UPPM based on knowledge graph is proposed.The model can automatically mine deep preferences of users by propagating user interest on knowledge graph.The use of attention network can adaptively distinguish the importance of user preference features to charactering the user's final preference vector at different propagation stages.UPPM combines feature-based methods,path-based methods and attention mechanism and applies them to personalized recommendation.The experimental results of click through rate prediction and Top-K recommendation tasks on real datasets show that the performance of UPPM model is significantly better than other methods commonly used in recommender systems.(2)To overcome the problem that the recommendation methods,including UPPM,focuses on modeling user-end or item-end knowledge graph,this paper proposes an end-toend user-item knowledge joint representing model UIKJR for personalized recommendation.UIKJR consists of two parts: item feature modeling and user personalized preference modeling.Item feature modeling task learns item feature representation by aggregating neighbor information of entity.User personalized preference modeling task can mine user's personalized potential preference by propagating user's interest to neighbor entity.UIKJR models user preferences and item features jointly,which can map user preferences and item features to the same vector space,and help to explore user latent preferences more comprehensively,so that the recommendation results are more rich and novel.Experimental on real datasets show that the performance of UIKJR is significantly improved compared with UPPM and other commonly used comparison methods.