Research And Implementation Of Recommender System Based On Knowledge Graph

With the rapid development of the Internet information age,the problem of information overload has become increasingly prominent.Recommender systems,an effective way to alleviate the problem of information overload,emerge as The Times require.Under the dual stimulus of users' demand for the accuracy of personalized recommendation and the demands of markets for the traffic monetizing,how to further improve the accuracy of traditional personalized recommender systems has become one of the research hotspots in both academia and industry.Knowledge graph,a form of information organization describing entity relationships in the objective world,can effectively enhance the semantic information of data and generate personalized recommendations more accurately.However,current knowledge graph based recommender systems still have many problems:1)Knowledge graph is general domain oriented.There is a problem of heterogeneous representations between knowledge graph and the users'interaction history in the recommendation scenario.2)Existing mainstream candidate generation algorithms lack of relevant research combining with knowledge graph.3)Existing knowledge graph based recommendation ranking algorithms ignore the modeling of high-order combinatorial features among different neighbor information on the knowledge graph.4)In the field of the technological services and resources,the practice and implementation experience of recommender system in relevant platforms is still insufficient.To solve the above problems,this paper focuses on the effective application of knowledge graph in recommendation scenarios and the implementation of knowledge graph based recommender systems.The main research work includes:1)Proposed a compressed collaborative knowledge graph and a dual graph transformer network based on the compressed collaborative knowledge graph mentioned above for candidate generation.By establishing the connection between users and the neighbors on the knowledge graph,the representations of users and items is modeled,which are used for candidate generation.In the public datasets Amazon-book and Yelp2018,we improved the Recall@20 by 4.30%and 9.25%respectively over the state-of-the-art.2)Proposed knowledge graph based wide&deep framework for recommendation ranking.The framework implements the modeling and integration of the neighbor information on the knowledge graph from the width and depth dimensions,which helps realize the personalized recommendation.In public datasets Amazon-book,Yelp2018 and Last-FM,we improved the F1-score by 8.59%,14.36%and 15.22%respectively over the state-of-the-art.3)Designed and implemented a knowledge graph based recommender system of the technological services and resources,which includes the candidate generation model and ranking model proposed in this paper.The recommender system generates personalized recommendations accurately based on the users' interaction historical behaviors and the knowledge graph of technological services and resources.The effectiveness of the algorithms proposed in this paper is verified in the actual application scenarios.