Research On Deep Recommendation Algorithm Based On Review Text

With the rapid development of the Internet,the amount of data that people can access has increased greatly,which makes it difficult for people to capture what they need from the massive data.The emergence of recommendation system can solve this problem well,and can meet the needs of users for personalized,so in recent years,it has been widely concerned by researchers.Traditional recommendation algorithms extract information only based on users’historical rating data,but this method faces serious problems of data sparsity and cold start.With the increasing abundance of information generated by Internet users,the integration of auxiliary information of users and items is an effective way to solve sparsity and cold start.Auxiliary information includes review text,social network,multimedia information(such as images and audio),etc.Review text,as the direct feedback of users on the item,contains information such as user preferences and item attributes.Aiming at how to effectively combine reviews with this paper to improve the accuracy of recommendation,the research content of this paper is as follows:(1)this thesis proposes a recommendation algorithm based on comment information(CRPS).Firstly,the dynamic word vector model BERT is used to process the review text to solve the problem of word ambiguity.Considering the different styles of each user’s reviews,multiple review vectors are cross-fused through multi-head attention network to obtain the review vector that integrates the overall style.Get user and item review characteristics by aggregating each review vector of the user and item.Finally,the review features of users and items are embedded and spliced with the ID of users and items through deep cross network to realize the high-order cross features,and the score prediction is carried out through fully connected layer.Through experiments on Amazon data set,compared with other review-based recommendation models such as DAML,the algorithm has better recommendation performance.(2)At present,most recommendations based on graph neural networks only use the ids of interactive objects to extract features,ignoring the features of users and items implied in other information.Therefore,this paper proposes a hybrid recommendation algorithm(GraphCrps)that integrates node features and review features.Compared with the recommendation model based on graph neural network,the introduction of review information can obtain more complete user and item feature information,and alleviate the problem of sparse interactive data.The model uses graph neural network to extract node information in the interaction graph,and uses the same feature extraction method as the CRPS model to process the review information.(3)Although the GraphCrps model has a good recommendation performance,the independent modeling of review and interaction graph ignores the correlation between information,so the CrpsMf model is proposed in this paper.Compared with the CRPS model in structure,the model introduces the ID and score of the review object.Since reviews of different objects and ratings have different values,this method helps to mine more valuable comment information and optimize recommendation performance.Finally,the experiment showed that compared with GraphCrps model,CrpsMf improved the accuracy of recommendation.