Explainable Recommendation Based On Review Filtering

With the rapid development of Internet technology and the rapid growth of information resources,users are faced with severely overloaded information,making it very difficult to quickly find target information resources.In response to this problem,a recommendation system came into being,which can provide personalized recommendation services to users according to different user preferences.The recommendation system has been developed for a long time,and most of the current researches focus on the task of recommendation accuracy.However,in addition to satisfying the accuracy of recommendation,recent researchers have also paid attention to the field of interpretable recommendation.Interpretable recommendation means that when a product is recommended to a user,an explanation is given to explain why the product is recommended.Explainable recommendations will increase the credibility of the recommendation results,thereby increasing user satisfaction.In scenarios such as e-commerce,movies,books,etc.,users often comment on items,but most of the current research on recommendation systems focuses on the effective use of reviews,and reviews are an important way to reflect user preferences.Therefore,finding an appropriate way to use user comments can improve the accuracy of the recommendation system and the effectiveness of interpretation.Aiming at the two issues of recommendation accuracy and interpretation effectiveness of the recommendation system,the thesis establishes a multi-task learning framework to complete end-to-end learning,which can simultaneously obtain the user's recommendation score and recommendation reason for the current product.In order to use the cross information of the two tasks and improve the effect of the two tasks at the same time,the RFCAML model is proposed.The historical comment information of users and products is introduced into the model,and sentiment analysis tools and classification algorithms are used to sort and filter the comments by importance.Use the collaborative attention mechanism selector to select the filtered user and product review sets to obtain a selected review subset of user product pairs,and use Microsoft's concept map for each review in the final selected review sub-set Carry out concept extraction to obtain the key emotional feature words of the selected reviews,and again apply the collaborative attention mechanism selector to select the concept words,and use the selected comments and key concept words of the user or product to represent the embedded vector of the user or product,and The embedding vectors are input into the network of recommendation score and explanation generation tasks respectively,aiming to solve the two problems of recommendation accuracy and explanation validity of the recommendation system at the same time.This article first builds an interpretable recommendation multi-task model framework based on comment filtering,and preprocesses it on Amazon's public data set to meet the model data input format requirements.On this basis,we use regression common indicators and common machines for natural language processing.Translation indicators evaluate the model's recommendation accuracy and interpretation generation effectiveness.Experiments have proved that the combination of the comment filtering method proposed in this paper and the overall multi-task model framework based on the collaborative attention mechanism can improve the accuracy of recommendation and the effectiveness of generated explanation to a certain extent,so that the recommendation system can bring higher quality recommendations to users.Service and better user experience.