Deep Recommender System Based On Review Text

Traditional recommendation algorithms,such as collaborative filtering(CF),rely on user-item ratings.However,as the amount of data grows,the scoring data will become more sparse.This problem will affect the accuracy and interpretability of the recommender system.Nowadays,many e-commerce websites provide a lot of more semantic review text information which can alleviate problems of cold-start and data sparsity,and can make the recommendation results more interpretive.Therefore,many scholars are studying deep recommender system based on review text.However,how to use review text to express user preferences and item characteristics more accurately is the research focus of deep recommendation algorithms based on review text.This thesis presents a novel model called NAFSBG.The point is to more accurately express the user and item feature vectors,making the prediction score more accurate.Specifically,the innovations of this thesis are as follows:(1)We use the Sentence-BERT pre-training model to initially encode the review text.The traditional CNN is relatively weak in feature learning ability of review text,and cannot learn the feature representation of review text well.Sentence-BERT uses Transformer(the most mainstream feature extractor at the moment),so that the initial encoding vector of the review text is more accurate.(2)We design a double-line attention model to explore the usefulness of review text.This thesis shows that useless reviews are less valuable to users and will also damage the performance of the model.Therefore,it is necessary to use the attention mechanism to describe the weight of each review to express the characteristics of users and items.(3)We design a novel gating mechanism to effectively integrate the ID feature vector and the final feature vector of the review text.In the case of a large amount of data,compared to the past work,the use of the gating mechanism simply adds or splices the ID feature vector and the final feature vector of the review text,which can better integrate two features to obtain a more accurate overall feature representation.Comprehensive experiments on 10 benchmark datasets have demonstrated that NAFSBG outperforms state-of-the-art approaches such as traditional rating-based recommendation models and review-based recommendation models in terms of MSE evaluation metrics.And the higher the data set density(the more review texts),the more obvious the improvement of NAFSBG.