Application Of Bi-Direcitonal GRU Network Based On Attention Mechanism In User Rating Prediction

In recent years,the development of e-commerce is getting better and better,and has profoundly changed people's life style.Online shopping has become an indispensable part of people's life,but it also brings a lot of troubles.Now the e-commerce website is full of too many goods,users need to spend a lot of time and energy to find the goods they want.How to provide customers with the goods they like quickly and accurately has become a problem that must be solved,so a recommendation system has emerged.The recommendation system can learn the user's preferences and needs according to the user's historical data,such as historical browsing records,shopping cart records,favorites records,order records and other information,and then recommend products for them.The common algorithm in the recommendation system is collaborative filtering algorithm,which depends on the "user rating matrix".However,there are a large number of goods on the e-commerce website,and users will only buy a small part of them,resulting in the "user rating matrix" is very sparse,so the recommendation effect is not reasonable.The common matrix factorization method in collaborative filtering is to learn hidden factor vectors from the scoring matrix to represent users and projects respectively,which can alleviate the problem of sparsity to some extent,but also lead to the problem of opaque recommendation.With the increasing participation of users on the e-commerce platform,a large number of user generated content(UGC)have been generated,such as user's social information,implicit user feedback,emotion of comment text,text content of user's comments on entities,etc.These user generated content contains a lot of valuable information,such as suggestions for products and services,personal preferences and so on,so these information can become a reliable source of information for recommendation system.Aiming at the problem of sparse scoring data and incomprehensible recommendation results,this paper proposes a user rating prediction method:Bi-directional GRU network based on attention mechanism.It uses deep learning technology to mine hidden information in user review text and learn with user-item rating matrix to alleviate the sparse problem of rating matrix and increase the accuracy of rating prediction.The user rating prediction model proposed in this paper mainly includes three aspects:Firstly,two bi-directional GRU network based on attention mechanism is created to model user review documents and item review documents respectively,from which user text features and item text features are extracted.Then,the extracted text features are introduced into the probabilistic matrix factorization model as a priori mean of the hidden factors of users and items,so as to regularize the probabilistic matrix decomposition model,so that it can perform well on invisible test data sets.Finally,a new calculation framework for optimizing model parameters is proposed,which is different from the previous model in that all parameters are trained and optimized simultaneously by establishing a total loss function.The method adopted in this paper is to update each parameter iteratively in a specific order.When one parameter is updated,the other parameters are fixed as a constant.The optimization method can alleviate the problem that the model training is affected by the correlation of parameters.The model is validated by three subclasses in Amazon's review data set.Through the research work of this paper,the following conclusions are drawn at last:(1)The total number of users and items on e-commerce platform is huge,but the total number of goods purchased by each user is relatively small,which leads to the sparse user-item rating matrix.By analyzing the length of review text in data set,the length of review text in most items is more than 50,which shows that review text can alleviate the sparsity of rating data,so it can be integrated into the rating data for research and analysis.(2)By using deep learning technology to process review text,more accurate and comprehensive text features can be obtained.Because bi-directional recurrent neural network can capture the semantic and contextual information in the sequence,attention mechanism can acquire information with high relevance to a particular topic while ignoring less important information,so it can process longer text sequences.(3)The optimization method of updating each group of parameters iteratively in a specific order is adopted.When updating a group of parameters,all other groups of parameters are fixed to a constant value.This iterative training method can alleviate the dependence between parameters and reduce the training time of the model..