Research On Personalized Recommendation System Based On Review Text

With the prosperity and development of Internet online services,a large amount of data on the Internet increases the cost of people’s choice of information.The personalized recommendation system allows users to reduce decision-making costs and find items that meet their interests in the face of massive data.It can also help product providers to provide better services.However,most recommender systems today face the challenges of sparse interaction,long tail effect,and lack of recommendation interpretability.The comments written by users on items contain rich information about user preferences and item attributes.The full use of these implicit feedback interactive information can effectively solve the problem of data sparseness.With the advance performance of deep learning in semantic understanding,many studies extract relevant information from reviews for recommendation system.However,most of these methods do not jointly utilize explicit feedback and implicit feedback information contained in user behavior.This paper proposes two neural network models that integrate review texts into the collaborative filtering algorithm framework and an interpretable recommendation model based on text generation which proves the effectiveness of text on recommendation improvement.The main work of this paper includes:1.In the framework of recommendation algorithms based on matrix factorization,traditional methods only use rating interaction information,which is challenged by the problem of data sparsity.Recently,deep learning-based methods model static user preferences and item attributes from user and item reviews.This paper focuses on the personalization ability and dynamic modeling ability of recommender systems,and proposes a description-aware personalized recommendation method.In the word level,a personalized attention network is designed to select most informational words;at the review level,a description-aware cross-attention networks are designed to dynamically learn user and item representations.Extensive experiments on 6 Amazon datasets of different scales show that the proposed method can improve the accuracy of recommendation score prediction.2.In the autoencoder-based collaborative filtering algorithm framework,most users only interact with a small number of popular items,which leads to serious popularity bias.Considering that the semantics in the review text are rich in user preference features,this paper proposes a hybrid recommendation model that fuses the review text to the autoencoder.By designing an attention neural network,the review is used as a content signal to fuse with the collaborative filtering signal of the autoencoder.Extensive experiments on three real datasets demonstrate that the proposed method can significantly improve the top-N recommendation performance.3.Rating prediction for recommendation requires modeling high-quality user and item representations,and personalized text generation requires user and item representations as a condition for personalized generation.At present,most researches focus on a single task.In this paper,the two tasks of score prediction and text generation are integrated to form an interpretable recommendation system.After the user and item representations learned by the recommendation module are encoded by the variational autoencoder,they are input into the text decoder.Focus on solving encoding posterior collapse and decoder efficiency problem.Experiments show that sample from code is helpful for diversity text generation,meanwhile,lightweight decoders can also produce better text quality.The proposed recommendation model incorporating reviews achieves some improvements in rating prediction,top-N recommendation,and interpretable recommendation,respectively.