Research On The Rating Prediction Based On Dynamic Topic Analysis Of User Reviews

With the popularization of social commerce,users will continue to generate information,such as reviews and ratings that reflect product characteristics and user preferences in the process of using e-commerce.Integrating the review information in each e-commerce platform and using these review information for rating prediction is a research hotspot in the field of intelligent recommendation.In order to achieve cross-platform rating prediction and improve the accuracy of rating prediction,the paper has done the following three aspects.A method for mining hidden information of user reviews based on time window is designed.Under different time windows,the word contained in the user reviews is the network node,and the number of word co-occurrences is used as the edge weight of the connected nodes,these construct a network model of user review information.For the same product,a dynamic user review information network is built on the target platform and the auxiliary platform respectively.The two information networks are connected by the same network node to form a dynamic cross-platform information network model.In the model,the method of calculating the degree of association and the similarity of user reviews are designed.Using the degree of association and centrality to mine the hidden information of user reviews on the target platform,the hidden information of reviews from target platform to the auxiliary platform.Combining the hidden information of the review with the original commentary information can more accurately describe the user's true feelings about the product,and reduce the error of the rating prediction.Cross-platform rating prediction can be achieved through the hidden information of user review form target platform to auxiliary platform.A review-preference dynamic mapping method based on time window was designed.On the basis of the dynamic topic model,analyzing the user's reviews dynamically,and mining the potential change rules of the topic words in different time windows.The user's preference evolution of the product attributes is characterized by the change of the probability value of the topic words,so that the user's rating prediction has timeliness.Under the specified time window,merging the similarity and mutual information between the topic words,and establishing the dynamic hierarchical tree of the topic words,so that the level of the topic words can dynamically represent the degree of influence of the topic words on the user's rating.Finally,proposing a user preference vector generation methodbased on the dynamic hierarchical tree of topic words.The user reviews are mapped in the hierarchical tree under each topic under a specified time window,to generate a user's preference vector.It show that user reviews under different time windows are mapped into the vector space of the same dimension.Based on the dynamic cross-platform information network model,the rating prediction based on the dynamic hierarchical tree of topic words further reduces the prediction error.In the process of rating prediction based on user preference vector,a two-stage optimization method for GBDT-LR prediction algorithm is proposed.In the first stage,the GBDT-MCLR(GBDT-MultiClass LR)prediction model is generated.GBDT-MCLR incorporates the idea of clustering into GBDT-LR,and proposes the ACK-Means(Adaptive Canopy+K-Means)clustering algorithm.The ACK-Means clustering algorithm can automatically select better cluster numbers and cluster centers according to the preference vectors,and divide the user preference vector into each class.GBDT-MCLR generate a fitting function in each class according to user preference vectors belonging to the class.The predicted value is calculated based on the corresponding fitting function.In the second stage,GABC-MCLR(GBDT and Binary Classification-MultiClass LR)prediction model is generated.GABC-MCLR converts the calculation process of prediction value into a quadratic equation solving process,and selects a quadratic equation according to the binary classification algorithm.The optimal solution is the new predicted value,and the new predicted value is closer to the true value than the original predicted value.The paper uses the user reviews on the e-commerce platform amazon and eBay as the data set.The three parts of the paper are validated by experiments.And each part is based on the previous part of the work,the prediction effect of user ratings is further improved.