Design And Implementation Of Personalized News Recommendation System Based On Deep Learning

In order to meet users' demand for personalized news browsing and adapt to the increasing number of news,personalized news recommendation system becomes more and more important.When personalized news recommendation is carried out,it is generally necessary to extract user interest and model the user to recommend news content according to users' interests.However,when modeling the user,the user's recent behavior is usually sampled to express the user's interest.There will be problems when the user's behavior sequence length is short or long.In order to solve these problems,this thesis includes the research of personalized news recommendation model and the design and implementation of personalized news recommendation system.In the research of personalized news recommendation model based on deep learning,this thesis considers from the perspective of data,and proposes a News CTR Prediction Model Based on User Behavior Fill and Select Model(UBFS).In the UBFS,the algorithm of user behavior sequence filling is proposed.When the user behavior sequence is short,part of news that user likely interest and potentially click on is selected to fill.Then,the thesis proposes the selecting algorithm of user behavior sequence.When the user behavior sequence is long,part of the user behavior sequence is selected to represent user interest,and the relationship between each behavior and candidate news is considered in the selecting process.Pretraining model Sentence-Bert is used for news modeling and user modeling in filling and selecting.Finally,the research of news click rate prediction is conducted.According to the behavior sequence of end users and candidate news,the click probability of users on candidate news is obtained,so as to indicate the degree of users' interest in candidate news.In this process,the pre-training model Bert is used to model news and user.Compared with other recommendation algorithms in the field of news,the UBFS algorithm proposed in this thesis improves by4.1% on AUC,8.9% on MRR and 9.7%(8.2%)on NDCG@5(NDCG@10).In the aspect of personalized news recommendation system design and implementation,this thesis designs and implements personalized news recommendation system based on deep learning on the basis of UBFS research.The system architecture includes: data layer,policy layer,feedback layer,application layer.In this system architecture,each layer structure is designed and implemented,especially in the strategy layer design and implementation,the news recall algorithm based on news semantics is used to recall news from the news list quickly and efficiently,and UBFS is used to sort candidate news,making the sorting results more accurate and reliable.To solve the problem of users' cold start,recommend news to users comprehensively considering multiple aspects,so that the recommendation results can be diversified and accurate;In the final sorting process,news recommendation sorting is carried out from multiple perspectives: accuracy,freshness and popularity.The recommendation results can recommend the news with high accuracy,freshness and popularity according to the user's interest.According to the above system design and implementation,the system function test and performance test.Functional test shows that each module in the system has complete function to meet functional requirements;Performance test indicates that the system can meet performance requirements from response efficiency,high scalability,reliability,and maintainability.