A New Bayesian Recommendation Algorithm On Fusing User Ratings And Reviews

With the rapid development of Internet,humans have entered a new era of intelligence.Due to the exponential growth of the amount of information and gradually increasing types of data,traditional recommendation system can not meet the requirements of both merchants and consumers.The key of the recommendation algorithm lies in how to find out more preference information from users.Then certain recommendation can be given to users according to their interests.Besides the rating that can reflect the users' preference information,the reviews written at the time of rating are also included.Tons of users' preference information and potential characteristic attributes of goods can be discovered by making semantic analysis about reviews,which can improve the problems of data sparseness and cold start.The research of the thesis focuses on the rating prediction of recommendation system.The abstract users' preference is integrated into the LDA(Latent Dirichlet Allocation)subject model in the thesis,and the Bayesian rating prediction model of users' rating and reviews is introduced.Through the experiments based on the public data set of Internet,it has been proved that the theoretical method put forward in the thesis can significantly improve the accuracy of rating prediction.There are two main points about the major tasks and contributions in the thesis.(1)LDA document generation model will be brought into the collaborative filtering recommendation algorithm in the thesis.The abstract users' preference will be explained by subjects and then the rating generation model is put forward on the basis of Bayesian Theory.The model adopts Gibbs sampling algorithm to make parameter estimations about the sample information observed.Regarding to the given users and goods,the model can predict the rating of goods from the users.The experiments has been conducted with the MovieLens data sets,the results show that the rating prediction algorithm put forward in the thesis has higher accuracy than traditional recommendation algorithm like Item-CF,SVD++ and PLSA-CF.(2)In order to find out more preference information from users,semantic analysis about the reviews at the time of rating and the simulation of generation process for each word in the reviews are made in the thesis.The thesis puts forward the Bayesian rating prediction model integrated with users' rating and reviews.A text sentiment analysis model is put forward on the basis of LSTM,which is used to make sentiment analysis of reviews.The problem of mismatch between ratings and reviews is solved.(3)The experiments has been conducted with the Jingdong's 23 data sets,which has proved that the rating prediction accuracy of the algorithm improves significantly.At last,the thesis demonstrates that the recommendation algorithm has better interpretability and can improve the problems of cold start of recommendation system and data sparseness.