Research On Recommendation System Based On User Reviews

With the development of the Internet,the number of users and commodities on e-commerce platforms has increased dramatically.The needs of users vary from person to person.Recommender system can help users find their favorite products from a large number of products,which not only improves the user's shopping experience,but also brings huge economic benefits.The traditional recommender system takes user review as input information,resulting in the deficiency of input data.To solve the problem of data deficiency,auxiliary information is introduced as input information.User review text is informative which makes up for the problem of data shortage.Compared with traditional Recommendation system algorithm,the recommendation algorithm based on deep learning can directly extract features from the data base.With its strong anti-noise ability,it can learn the characteristics of users and products more accurately.In recent years,it is widely used in various kinds of recommendation tasks.But this algorithm still has some shortcomings.First,the recommendation model based on convolutional neural network can only extract the features of fixed receptive field,rather than all the features.Second,the current recommendation algorithms based on deep learning usually use convolutional neural network or recurrent neural network to extract the features of reviews,but it cannot fully express the feature information of users and products.Third,this algorithm does not take into account that the importance of reviews and words is different to user and product modeling.This article proposes corresponding solutions to the above problems.The main research contents are as follows:Firstly,given that convolutional neural network can only capture the word information of a single receptive field when extracting review features,this research uses gating mechanism to adaptively extract review features extracted by multiple convolutional neural networks;Secondly,since using the convolutional neural networks or recurrent neural networks alone cannot extract local and global information of reviews comprehensively,this research uses adaptive convolutional neural network to extract the local features of reviews and uses gate recurrent unit(GRU)or transformer to extract the global features to fuse local and global features and better express comments information;Thirdly,in order to better evaluate the importance of user's reviews on the features of user and product modeling,this research uses multi-level attention mechanism to evaluate the usefulness of reviews to measure the impact of different reviews and words on modeling;In summary,this research constructs an end-to-end,deep-learning based scoring prediction and recommendation model through feature extraction of user rating and reviews.Compared with the baseline model,the effect on five Amazon product review datasets is improved by 2.93%-6.26%.The experimental results show that the proposed model is reasonable.