A VAE-based Review Generation Style Interpretable Recommender System

With the rapid development of the Internet,the continuous accumulation of online data has led to the problem of information overload,making it difficult for users to quickly and accurately identify valid information for effective decision-making.In this case,recommendation system,which actively recommend information by learning each user's interests from her historical information,has become one of the most popular techniques to resolve the information overload problem.An effective personalized recommender can improve users' online experience and thus obtain huge economic benefits for the companies.Therefore,to the topic that how to build more accurate and effective recommender systems has attracted increasing attention from academia and industry.A explainable recommendation system refers to that the system not only generates personalized recommendations for target users,but also gives corresponding explanations for the recommendation.It can increase the transparency of the system,allow users to better understand and accept the recommended items,and correspondingly promote the overall benefit.On the other hand,online reviews potentially reflect valuable information such as user preferences and item characteristics,and are thus excellent valuable information sources for recommendation explanation.Previous works that make use of review text data to achieve recommendation explainability can be conclude into two categories: retrievalbased methods and generation-based methods.The former ones often filter the words or fragments from the review text,and add it to preset mode(such as template sentences)as recommended explanation to the user.The latter,often put both prediction task and explanation task on recommendation into a multi-task learning framework,makes recommendation list and generates sentences as corresponding explanation simultaneously.Previous models can work normally when the review text is available,but for a new user,the model will be difficult to function properly to generate an effective or reasonable textual explanation if the review is missing or of low quality.Aiming at providing accurate recommendations for users,and making the textual explanation generation work when the high quality of reviews are unavailable,this paper proposes a novel deep learning-based recommendation model,which is the Variational Autoencoder-based Review-Style Explanation for Recommendation(VAE-RSE).VAE-RSE completes both the prediction task and the explanation task in an end-to-end manner.The model quickly generates a recommendation list based on a few of interaction data,and can generate corresponding explanation texts to users by utilizing the user feature hidden in interaction data when reviews are unavailable.VAERSE consists of four functional modules: Topic Extractor,Text Generator,Predictor,Transformer.The first three modules all set as variants of the variational autoencoder(VAE),and the last module is a simple conditional adversarial network(CGAN).The Topic Extractor extracts the topic vector from reviews for regularizing or supporting other modules.The Text Generator learns the text generation mode through the autoencoding on reviews.The Predictor accepts user interaction information and the corresponding item topic vectors,to generates a recommendation list.The Transformer acts as a link between the Predictor and the Text Generator,making it possible of multimodal spanned comment generation in the absence of comments.In order to verify the effectiveness of VAE-RSE,this paper conducts substantial experiments on the public datasets,i.e.,Amazon review data.Experiments show that in the top-k recommendation task,in terms of Recall,Precision,and NDCG,VAE-RSE significantly outperform state-of-the-art methods.In the task of evaluating the quality of generated explanation,we used both metric ROUGE and manual evaluation methods.The results show that VAE-RSE can generate explanations that are rich in content and consistent with the semantic information of real reviews.Moreover,the model can still output reasonable textual explanations for users whose reviews are missing.In addition,the ablation study of VAE-RSE is performed to verify the effectiveness of each part of the model.