Research Of Rating Prediction Based On Review Text

With the flourishing development of the Internet,network information is growing rapidly.High-speed flow of information can bring convenience to people's daily life,but it also means that it is more difficult for users to find the information they need quickly and accurately.Once the user opens the page,clicks the mouse,fills out the form and perform other operations can generate data.The recommendation system can infer the user's interest preferences from these data,thus can recommend personalized information to users.The recommendation system has practical application value.For example,Taobao,Headlines Today and NetEase cloud music have gained an advantage in the similar market competition with their accurate recommendation.At the same time,the recommendation system is also a hot topic in the academia.There are two main research directions of recommendation system,one is Top-N recommendation for users,the other is to predict users'ratings on products.The Top-N recommendation is to recommend a list of products for the user,the essence of which is the rating prediction.In recent years,more and more researchers have paid attention to the recommendation algorithms of rating prediction based on review text.The review texts written by users contain more reliable information,from which the attribute characteristics of products and the users'interest preferences can be deeply mined to achieve more accurate recommendation.This paper is based on the review text to solve the common rating prediction problem at present.The main work is as follows:(1)We propose a recommendation model based on review graphs.The model includes three stages:knowledge processing,feature construction and predictive regression.Firstly,the Doc2vec model is used to extract the semantic features of the reviews.Then,the method based on graph model is used to extract the keywords from reviews,and the knowledge graph is used to construct the semantic relationship between the keywords from reviews to expand the user's interest preferences.The feature of users and products are represented by the fusion of the semantic features of the reviews and the features of the network nodes extracted from the knowledge graph.Finally,the libFM model is used to predict the ratings in the predictive regression stage.(2)We propose a rating prediction recommendation model based on dual-level attention mechanism.This model contains two parallel convolutional neural networks to jointly learn the hidden features of users and products.In modeling phase,both fine-grained words and coarse-grained reviews are taken into account.The connected word vectors and review vectors are used as input of the network,and a semantic first-order jumping method based on Word2vec is used to represent the review vectors,which further enriches the semantic expression of the reviews.A attention layer is designed before the convolutional layers to reinforce the contribution of the important features to the rating prediction,and to increase the interpretability of the model.The top layer is motivated by factorization machines to simulate the interaction of higher-order latent features for the rating prediction.Experimental results show that the proposed method has lower root mean square error than the benchmark methods,thereby effectively improving the accuracy of the rating prediction.