Research On Recommendation Algorithm Based On User Review

With the fast development of the Internet,the information on the website is expanding rapidly,which causes the information overload,even information disaster.In the face of huge amounts of information,people tend to consume a lot of time when they get the information that they are interested in,even are lost in the sea of information.The personalized recommendation is effective to relieve the information overload and plays an important role in many fields.However,traditional recommendation technologies only rely on the ratings which given by users or users' browsing behaviors to analyze the users' preference.They don't make full use of the users' reviews and the recommendation results are slightly inaccurate.In order to improve the user experience,the valuable information is extracted from the reviews with the product review mining technology,so as to realize the active recommendation to the users.In this thesis,the user interest model is constructed by extracting the attribute words of the comment text and then an improved matrix factorization recommendation algorithm based on SVD++ is proposed.The movie recommendation is as an example in the thesis and the main research work are as follows:(1)Extracting the attributes of comment texts.According to the characteristics of the movie reviews,extracting the feature words of comment texts to achieve the purpose of obtaining the attribute characteristics of reviews.The existing feature extraction techniques based on statistics don't fully take the topic and semantic information of the document into account.In view of this limitation,a new method of extracting feature words from a comment text is designed,based on word vector and clustering techniques.Taking into account the field of movie reviews,a domain related corpus is established first.Then the word vector technology is applied to the text feature extraction and obtaining the attribute words through clustering.The experiment results show that it's feasible to extract the attribute words of the movie reviews and the domain specific corpus built by self are available to improve the effect of feature words extraction.(2)Constructing user model based on reviews.The user model representation methods which are widely used at present are discussed,and then a user interest modeling method based on attribute characteristics of reviews is proposed.Firstly,the weights of the attribute words in the reviews are calculated,and then the user attribute categories are represented to complete the user interest modeling.The experimental results indicate that the method can effectively represent the users' attention on the various feature levels of the product and realize the representation of users' interest.(3)Improving the matrix factorization model.In the traditional collaborative filtering recommendation algorithms,data sparseness is a serious problem.Thus a matrix factorization model based on the fusion of comment information and rating matrix is put forward.The model uses the users' preference at characteristic levels of the product to obtain the preference similarity among the users.Then the preference similarity is used as a correction term to improve the SVD++ model.Through the comparison experiments,it's proved that the improved algorithm which combines users' review information and rating information has high prediction accuracy.(4)Implementing the movie recommendation prototype system.First a recommendation engine is designed by the steps of processing data,computing matrixes,generating recommendation lists and presenting recommendation results.A movie recommendation prototype system based on JavaScript and CSS is implemented finally.