Research On Generative Model Based On Review Sentiment And Recommender System

After the users consume on the e-commerce websites,the shopping feedbacks are included.And the feedbacks include the numerical scores and the text comments.Scores have always played an important role as predictive targets in recommender systems.The other part of the feedback,text comments have not received the same attention,until in recent years,researchers began to research on the combination of comments and scores.The amount of information in a comment text is several times more than the users' other behavioral data(such as purchasing,rating,clicking,browsing,etc.).Analyzing the comment text is an important way to improve the recommendation effect.The users' comments will include several aspects of the information,including the shopping experience,the characteristics of the article,the users' own characteristics,etc.The information can help researchers to quickly and accurately create user profiles,item characteristics vectors and other data models,and recommend accordingly.Among them,the most relevant to the user's rating is the user's shopping experience,and the user may show some emotional tendency in the comment.And these emotional tendencies are closely related to the score.Negative comments tend to give users a lower rating.There are still some problems in the existing research on e-commerce user reviews.First,these studies focus on aspects or features,require extensive manual labeling,and require repetitive work if the item type changes.Second,the researchers only considered the characteristics of the comment while ignoring the user's emotional inclination in the comment.Third,some studies used LDA for thematic analysis,and the topics they obtained were less interpretable.Fourthly,some researches use maximum likelihood method to solve the probability model,fall into the local optimal solution easily and rely too much on the initial parameters.This paper presents a recommendation system based on a review text model that analyzes the users and items in comments and uses them to make recommendations.The research results of this paper are summarized as follows:(1)A method for preprocessing comments is proposed in this paper.When the user needs to predict the item's rating,all the user's previous reviews in the system are aggregated into user comments,all the reviews obtained before the item are combined into reviews,and then the two newly added comment texts are analyzed.In this paper,we analyze the emotions in the text and mark the emotional tags for each text according to the text sentiment analysis algorithm,and generate the emotional topic-probability distribution of words by TFIDF method.(2)In this paper,a review text generation model is proposed.Words in the comments are generated from the background thesaurus the emotional thesaurus or the thematic thesaurus.This article looks at the information,features,and emotions in the commentary from two perspectives,each of which affects the generation of the final commentary.By introducing the background thesaurus to reduce the interference caused by common words.This paper uses Gibbs sampling to estimate the parameters in the model.(3)In this paper,the real data set is used to validate the prediction accuracy of the model.Compared with other recommended algorithms,RMSE is used as the evaluation index.Experimental data show that the proposed model proposed in this paper has a significant improvement over other algorithms.(4)Based on the review-based recommendation system proposed in this paper,a prototype recommendation system based on e-commerce website is designed and implemented.The main functions of the system include the user's registration,login,shopping and feedback,the system will collect and analyze the user's behavior data and feedback comments,personalized recommendation for the user.System uses a layered architecture,divided into display layer,business layer,data layer,storage layer and offline computing layer.This paper firstly introduces the research significance of the recommendation system based on comments,introduces the current situation and existing problems of using the user comments research,and puts forward the technical route of this research.Then the model proposed in this paper is described in detail.Including the parameters of the model,the probability map model and the parameter estimation method,the experimental verification model accuracy.Finally,a prototype recommendation system was designed and implemented to show the application of the model in the real world.