| The e-commerce platform provides users with a variety of feedback methods,such as reviews,tags,and ratings.As a supplement to ratings,reviews contain more fine-grained feedback information.Products have a lot of attribute characteristics,and users specifically explain their preference for each attribute of the product in the reviews and their reasons.Therefore,we can understand the distribution of interest weights of users through reviews,and discover the main factors that affect the purchase of products.For users with historical purchase records,the recommendation algorithm based on collaborative filtering only uses ratings to recommend products.If the rating and review information can be used at the same time,the prediction bias of the rating information can be corrected.This paper proposes a review aware hybrid recommendation algorithm SATRec.On the one hand,SATRec mines the collaborative filtering features of users and products from the rating matrix.On the other hand,all reviews are input into an encoder network to obtain preencoded review vectors.Then Convolutional Neural Networks(CNN)are used to extract the contextual features of users and products from reviews.Finally,SATRec uses an attention mechanism to dynamically combine rating-based collaborative filtering features and reviewbased content features to make rating predictions.The experimental results on the public recommendation data set Amazon show that the algorithm can effectively improve the recommendation effect,and the attention network is the key to the efficient combination of features.Fully-cold-start users are those who do not have any historical reviews and rating information in the domain.For fully-cold-start users in a specific domain(target domain),they can use their ratings and review information in other domains(source domain)to alleviate the fully-cold-start problem.This paper proposes a general review aware cross-domain recommendation framework RACRec.First,all reviews are pre-encoded into a vector form by the encoder,and the adjacency matrix is used to dynamically select the most important reviews of a pair of user in the source domain and product in the target domain,then the product feature vector is extracted from the product review text in the target field through the encoder-decoder structure.Secondly,an encoder-decoder-based migration model is designed to extract the domain-specific feature vectors and the domain-shared feature vectors of the fully-cold-start user in the source domain.The domain-shared feature vectors will be combined with the product feature vectors in the target domain to obtain the predicted rating.This paper also explores the impact of the encoder-decoder structure of the user feature extraction part on the prediction effect.Experimental results in multiple domains on the Amazon data set show that the performance of the RACRec model based on DNN-DNN is superior to the latest comparison algorithm.Compared with DNN-DNN,the DNN-DNN structure with parameters can further improve the recommendation effect.CNN-LSTM can avoid the problem of limited feature extraction ability and information loss of DNN,and the effect is most significant. |
