Research And Application Of Review-based Rating Prediction Method For Deep Recommendation Model

The emergence of recommendation system alleviates the problem caused by information overload and provides accurate items for users.But the performance of existing recommendation models are limited from lots of problems such as data sparsity,cold start.With the rapid development of deep learning,text processing technology ushered in a new breakthrough.Researchers begin to use text processing technology to quantify the review text and use the information of the review text to implement recommendations,which alleviated the data sparsity to some extent.Under this background,the deep recommendation model based on review text has been greatly concerned.How to extract more valuable information from the review text and improve the performance of the recommendation model has become a hot topic.The thesis takes the review text as the main research object and discusses the review-based rating prediction method for deep recommendation model.Firstly,the new dual attention layer is integrated into the feather extraction ability of text processor optimization model based on convolutional neural network;secondly,the improved aspect recognizer based on fine-grained text and aspect sentiment analyzer are used to generate aspect sentiment information,and the sentiment information is applied to the original model to improve the score prediction effect;finally,the e-commence prototype system is designed and realized.The specific work is as follows:(1)Aiming at the problems that current mainstream review-based rating prediction method for deep recommendation model cannot give full consideration to the importance information of review text at different levels on the model effect,a new rating prediction method for deep recommendation model DATCo NN is proposed.The model is improved on the basis of mainstream text processor.On the one hand,the word level attention layer is integrated with context information in the process of word vector processing,and the convolutional neural network is used to strengthen the coupling between words and highlight the words with higher importance;on the other hand,the review level annotation with time factor is embedded in the process of integrating user and item vector matrix In the semantic layer,the concept of forgetting rule of human brain is introduced,and the time factor is integrated into the attention mechanism to fit the change of user's interest in the project with time in the actual recommendation scene,so as to reduce the impact of useless review on the modeling effect.The results show that the improved model can well mine the word level and review level information.Meanwhile,the improved model can extract features better.(2)It is an abstract process to extract the features of the review text by using the deep learning technology.It is impossible to extract the concrete information of the review object and the user's emotional tendency to the review object.The introduction of the external annotation aspect information into the model to define the aspect and emotional polarity of the review text can make up for the shortcomings of this aspect.However,how to accurately identify the aspect and how to apply the emotional information of aspect to the recommendation model are the difficulties of this scheme.In view of the above difficulties,based on the original model,an review-based rating prediction method for deep recommendation model ADATCo NN is proposed.On the one hand,the model divides the review text into clause segments according to punctuation and dependency syntax analysis results,and transfers the aspect to more suitable clause segments.The recognition device is trained by using the aspect information of the fine-grained text,which can alleviate the conflict of aspect categories caused by the different meanings of the segments before and after the review text.On the other hand,the user's emotional tendency is calculated according to the aspect information and the emotional polarity generated by the emotion analyzer ATAE,and the above information is integrated into the user preference vector and the item attribute vector.Meanwhile,the loss function structure of the recommendation model is optimized by using the loss of the review text in the aspect recognition process,and the aspect information and recommendation module are combined to enable the model to capture the attribute of the aspect and accurately the emotional orientation of the users.The results show that the improved ADATCo NN can integrate the specific user emotion information into the recommendation model better.Meanwhile,the improved model obtains prediction results more accurately.(3)On the basis of the above research,the thesis has designed and realized the ecommerce prototype system.The system adopts the current mainstream technology architecture,realizes the main functions of e-commerce platform,and encapsulates the recommendation model studied to provide product recommendation function for users,it shows the value of the recommendation model.