Research On Personalized Recommendation System Based On Personalized And Implicit Review Feature Acquisition

In the Internet era,with the development of recommendation systems,personalized recommendation systems based on reviews have gradually become a field of attention.Due to the development of e-commerce,people have become accustomed to posting their own reviews on items or information on Internet platforms.However,reviews contain rich user experience information and related attribute information of products,which leads to the research field of personalized recommendation system based on reviews.This field mainly extracts the personalized features of users and items by analyzing the review information of users and items,so as to alleviate the sparsity problem of traditional personalized recommendation systems and improve the performance of recommendation systems.However,existing review-based personalized recommendation methods generally adopt global text feature acquisition technology,ignoring that the understanding of the same review text will change with the specific users and products associated with it,that is,relying on a personalized association between a specific user and an item.On the other hand,current methods mainly focus on the information explicitly expressed in the reviews without fully mining the implicit information due to the relationship between users,that is,relying on the personalized association relationship between users.Therefore,this thesis first proposes a method for modeling personalized text feature acquisition associated with specific user and item pairs,and secondly proposes to use group common features as implicit features to study review-based recommendation tasks.The main contributions of this thesis are as follows:(1)Aiming at the problem that the understanding of the same review text will vary with the specific users and items associated with it,this thesis proposes a recommendation model to dynamically reconstruct multiple comments into a personalized document based on a specific user-product pair.For a given user-item pair,the model designs a constructor that uses a cross-attention mechanism to calculate and select important words for personalized document reconstruction.Second,the model designs a review encoder and a document encoder based on a cross-transformation mechanism to fuse user and item features on reviews and review documents.(2)Since the latent features of users are usually not expressive,in order to model the latent features of users hidden in groups,this thesis proposes a reviews-based recommendation model mining the common features of the most relevant groups as latent features.The model finds the most relevant groups through the collaborative filtering method,and designs a group common feature extraction module based on the similarity fusion between K member features.Secondly,the model also designs a review encoder to fuse the dominant features of users and items on reviews.(3)Extensive experiments on three real datasets demonstrate the effectiveness of our two models.In addition,this thesis also conducts experimental analysis on the core parts and key parameters of the two models.