Research On Collaborative Recommendation Algorithm Based On User Ratings And Reviews

With the advent of the age of big data and artificial intelligence,as an important technique to solve the problem of information overload,personalized recommendation system is concerned by more and more researchers.The traditional recommendation system based on collaborative filtering algorithms has achieved remarkable achievements and has been widely used.To complete the user's recommendation task,traditional collaborative filtering technologies mainly focus on the research of the user-item rating matrix and then make rating predictions based on the nearest neighbor model or matrix decomposition method.With the continuous expansion of users and items,the data sparsity and cold start-up problems of traditional collaborative filtering recommendation algorithms are gradually exposed and increasingly serious.The traditional collaborative filtering recommendation algorithms only consider user-item rating data,and ignore user reviews,user labels,item labels and so on.These text contexts contain abundant user preferences and item feature information,which are of great significance to improve the recommendation effect of the recommendation system.Therefore,based on the in-depth study of traditional collaborative filtering algorithms and user reviews analysis,two recommendation models based on user ratings and reviews are proposed.The main contributions are as follows:Firstly,a rating prediction model based on the review topic analysis is proposed.User reviews contain rich user and item feature information,while user rating data are the quantitative evaluations and expressions of user reviews.In this thesis,the LDA topic model is used to extract user preference vectors and item feature vectors from user reviews.Then,based on the regression model,the function relationship between the review feature vectors and the user ratings is established,and a rating prediction model based on the review topic analysis is constructed.The recommended tasks are completed according to the rating prediction results.Secondly,a recommendation model based on the review topics and latent factors is proposed.Based on the principle of matrix decomposition,the traditional latent factor model decomposes user ratings into implicit user preference vectors and item vectors.In this thesis,the LDA topic model is used to extract user preference vectors and item feature vectors from user reviews and then the vectors are embedded into the traditional latent factor model.Thus,a recommendation model based on the topics of reviews and latent factors is established.The recommended tasks are completed according to the rating prediction results.Finally,the two models proposed in this thesis are verified on the Amazon e-commerce review data sets and compared with traditional recommendation algorithms.The experimental results show that compared with the traditional collaborative filtering algorithms,the two models proposed in this thesis can effectively improve recommendation effects.Moreover,they have better interpretability and alleviate the data sparsity problem to some extent.