Research And Implementation Of Recommendation System Based On Social Network User Influences

With the rapid development of Internet social network platform,the rapid expansion of network information and resources,and the increasingly complex social relations of network users,the problem of information overload is becoming more and more prominent.As an information filtering system,the recommendation system constructs a personalized preference model by deeply mining and analyzing user historical behavior data,network user social relationships and website project features,and recommends the information needed by users.The text mainly studies how to use the social user's influence distribution and interest similarity to improve the accuracy of the probability decomposition model.Based on user activity,information quality and interest similarity,the local influence model based on PageRank is improved.The interest similarity calculation method is proposed based on user historical behavior and theme tendency.Finally,the two are merged into the probability matrix decomposition model,and the IS-PMF model is proposed.The main work of this dissertation is as follows:Firstly,this dissertation summarizes the background,significance and research status of the recommendation system and question-and-answer community,introduces several traditional algorithms of the recommendation system,text feature extraction technology and recommendation system related technology,and expounds the calculation and analysis methods of social network and influence,which provides a theoretical basis for the research of this dissertation.Secondly,this dissertation proposes to use user historical behavior data to comprehensively measure user-item scores.Because traditional recommendation algorithms often use explicit feedback scoring matrix for calculation,with the development of social networks,more and more social networking sites only have implicit feedback,and only using explicit feedback data will result in sparse data,cold start,etc.problem.So choose to use the user's historical behavior data to comprehensively measure the user-item score.Thirdly,this dissertation analyzes the user's historical behavior record and uses the LDA topic model to extract the topic vector of interest to the user.Finally,the similarity between users is obtained by comparing the relative entropies of the topic vectors between users.This is an improvement for the case where it is relatively difficult to measure user similarity in an implicit feedback scenario.Last,According to the characteristics of question-and-answer community,influence neighbors and similar neighbors are constructed according to improved local influence algorithm and user similarity calculation method.Then,based on the combination of user influence distribution and user similarity,it is integrated into the probability matrix factorization model.In the crawled network data set,the feasibility of the proposed recommendation algorithm is verified,and good experimental results are achieved.