Research On Recommendation Algorithm Based On Graph Neural Network And Review Semantics Learning

With the development of science and technology,it becomes easier and easier for people to obtain information.The network is full of a large amount of information to meet people’s needs,but also makes it difficult for people to quickly find the content they are interested in.Recommender system is an effective tool to solve this problem and has been successfully applied in many platforms.Recommendation algorithms based on collaborative filtering use the user’s past interaction records to learn the characteristic representation of users and items,and can match the most appropriate items to users.However,the previous methods that only use single interactive data such as rating are vulnerable to problems such as data sparsity and cold start.In many e-commerce websites,along with the ratings,there are text reviews,which contain rich semantic information.Many studies began to integrate the review information into the recommendation model to alleviate the above problems and further improve the accuracy of the recommendation model.However,the previous models have some limitations:(1)Review-based recommendation algorithms often use convolutional neural networks to extract review features.They can only learn the relationship between short-distance words,but can not learn the relationship between long-distance and discontinuous words.In addition,the existing methods do not fully consider the high-order collaborative relationship between users and items.(2)Sequential recommendation models often only use user interaction sequence data for learning,and the effect of fusing review information has not been fully studied.At the same time,the mainstream methods take the user as the center and models the user sequence,ignoring the collaborative relationship between the user and the item.This paper studies the above two problems,and the results are as follows:A recommendation algorithm GAR is proposed,which integrates review semantics and high-order interaction relationship.This method is a two view fusion method considering review view and graph view.In the review view,each review is constructed into a word co-occurrence graph,and the individual review embedding is obtained by using Gated Graph Convolution Network learning.The attention mechanism is used to weighted sum the review embedding to get the review representation of users and items.In the graph view,this method regards the interaction between users and items as the user-item interaction graph,embeds the review information into the edge,and uses the graph attention network to update the user and item nodes.Then,the method stacks multiple layers of networks to aggregate high-order neighbor information to obtain the high-order representation of users and items,and then integrates the representations of users and items learned from the two views into the prediction layer for rating prediction.The results on four public data sets show that GAR is better than the previous models.A sequential recommendation algorithm RGS based on review text pre-training is proposed.In order to better extract review features,this paper first designs a review encoder to mine review text features and output review embedding,then designs a pretraining task using the review to predict the rating to train the review encoder.Then,this paper constructs a dynamic interaction graph from the recent interaction sequence of users and items,embeds reviews as edge information,uses self-attention network and feedforward neural network to model the timing relationship of neighbor nodes,uses attention mechanism to aggregate the characteristics of neighbor nodes to update the center nodes,and iteratively obtains the high-order characteristics of users and items to predict users’ preferences for items.Experiments on three public data sets show that RGS is better than the previous sequential recommendation model.