Research Of Collaborative Filtering Recommendation Algorithm Based On Sentiment Analysis And Topic Model

In the era of Internet,personalized recommendation system came into being and developed rapidly.The core of the recommendation system is the recommendation algorithm,which is related to the user's satisfaction.Nowadays,the mainstream one is collaborative filtering(CF)algorithm which uses user similarity to calculate scores of one user to one product in the prefilling stage.In the prediction stage,the traditional collaborative filtering and topic model are combined to calculate the similarity of users with user topic distribution.But there are two problems.First,user preference information can't be expressed by a general score.Second,the combination of topic model and traditional collaborative filtering depends only on part of user ratings,which cannot catch the whole user scoring features.In order to further improving the accuracy of CF algorithm,we aiming at dealing with the above two problems,propose a prefill algorithm based on sentiment analysis and a CF recommendation algorithm based on topic model.In the prefilling stage,we use sentiment value to prefill original matrix according to the selection strategy.In the prediction stage,this algorithm introduces trust parameters on the original score matrix decomposition,and then use user factor vector in matrix decomposition to structure implicit topic model.The score matrix decomposition model and implicit topic model are combined to structure CF recommendation algorithm based on topic model.The algorithm trains the rating prediction model parameters on all users' scores,and gets more optimal models to improve the prediction accuracy by grasping the overall rating feature,so as to improve accuracy of CF algorithm.By the experiment and the test in the real data set,the influence of each parameter on the accuracy of the proposed algorithm is analyzed based on the control variable method.Compared with the user-based CF algorithm and the biased SVD algorithm,the results of the test verify that the proposed algorithm can enhance the accuracy of recommendation.