Research On Recommendation Algorithm Based On Review Text

With the development of the Internet,recommendation systems are widely used in various practical application scenarios,such as social media network applications,ecommerce platforms,and so on.The system has gradually become an indispensable and important tool for people by helping users to filter out the content of interest and reduce the cost of effective(user-interested)information acquisition for users.The higher the system recommendation accuracy,the better the user experience.Therefore,it is of great practical significance and value to study the recommendation algorithm to improve the accuracy of the recommendation system.However,recommender systems face many challenges,one of which is the sparseness of rating data.To mitigate the impact of this problem,scholars have considered making full use of other data sources.Therefore,the recommendation algorithm based on the review text has been developed and has become one of the hot research directions.The review text contains a lot of useful user information,and can represent the characteristics of items to a certain extent,so fully mining the rich semantic information in the review text can help the model to improve the modeling quality,thereby improving the ability of rating prediction.Existing recommendation algorithms based on review text have the following two limitations: First,most models do not fully mine review text information,resulting in insufficient feature extraction accuracy;Second,some models do not use scoring matrix,even models that use scoring matrix.Its learning of rating matrix information is still insufficient.Therefore,this paper conducts related research on the above two main problems and designs the corresponding model.The main work of this paper is as follows:1.A deep recommendation model RRMBA(Review Recommendation Model on BERT and Attention)based on review modeling is proposed.The purpose of this model is to improve the utilization of text information,thereby improving the shortcomings of the algorithm’s insufficient feature extraction ability.This paper first uses BERT with powerful feature extraction ability to model each comment,fully mines the information in each comment text,and obtains the text feature latent vector of each comment.Then,using the comment-level attention mechanism,the useful value is calculated for the latent vector of each comment text of users and items,so as to obtain more representative user preference features and item feature representations,and finally improve the prediction accuracy.2.A deep recommendation model DeepFTMM(Deep Fusion Reviews Text and Matrix Model),which integrates rating matrix and review text,is proposed to solve the problem of insufficient information mining of rating matrix.The model uses the current popular graph convolutional neural network to fully learn the higher-order implicit interaction between users and items,and then effectively integrates the user-item feature representation learned from the review text to obtain more accurate user and item features.Item representation,thereby reducing prediction error.For the two proposed models,this paper conducts experiments on five selected Amazon sub-data sets.Compared with the benchmark model based on deep learning,the average mean square error(MSE)of the RRMBA model is reduced by 3.53%,and the DeepFTMM model is reduced by 8.39%,which proves that the RRMBA and DeepFTMM models designed in this paper have improved the recommendation accuracy.At the same time,this paper also conducts ablation experiments of the corresponding modules for the two models to verify the rationality and effectiveness of each module.