Research On Collaborative Filtering Recommender System Based On Review Analysis

Recommender systems typically produce a list of recommendations to precisely predict the user's preference for the items.It is widely used in e-commerce websites,online content providers and social network and becomes one of the most important techniques to improve the service quality of these websites.In recent years,Collaborative Filtering(CF)is one of the most successful recommending methods.Using the ideas of collective intelligence,CF is able to produce precise recommendations for users.However,the data sparse problem and cold-start problem restrict the performance of traditional CF methods.In other words,when the interactive information of users and items are sparse,CF often becomes invalid.Recently,online user feedback accompanied with review texts has become increasingly common.It is noteworthy that the review texts often contain abundant information of users and items.This paper proposes two improved rating prediction methods by review texts analysis.Furthermore,the two methods improve the accuracy of rating prediction especially in cold-start situation.Firstly,this paper proposes a rating prediction framework based on distributed representation of document and regression model.The framework takes advantage of distributed representation of document to map the unstructured review texts into the same vector space,and furthermore constructs the feature vector of users and items.The framework trains several regression models to predict ratings.The framework is able to produce more precise features that characterize users and items and blends several regression models to improve the rating prediction accuracy compared with traditional CF methods.Secondly,because the review texts contain not only users' attention to the situation of the different aspects of items,but also users' sentiment for particular aspects of specific items,which always determines how users rate items.However,traditional latent factor models often ignore such review texts,and therefore fail to characterize users and items precisely.This paper proposes an extended Hidden Factors as Topics Model(HFT)(a model combining the Latent Factor model and the Latent Dirichlet Allocation)based on Aspect and Sentiments Unification Model(ASUM)(an extended topic model),called Ratings Are Sentiments(RAS).By combining users' sentiments in review texts and their ratings,the RAS model can learn more precise latent factors of users and items compared with the baseline models.The extensive experiments on large,real-world datasets demonstrate that the rating prediction framework and the RAS model perform better than both the latent factor model and the HFT model and alleviate the cold-start problem to some extent.