Research On Cross-domain Recommendation Method Based On Rating And Review Transfer Technology

With the development of the Internet,recommendation system has gradually become one of the essential applications in e-commerce,social networking,music,video and other fields.The traditional recommendation system is usually carried out in a single field,that is to say,items in the same field are recommended to users.When facing new users or the data is too sparse,accurate recommendation cannot be made.In recent years,cross-domain recommendation has gradually become a research hotspot in the field of recommendation due to the ability to use ratings and consider the influence of review information at the same time.Cross-domain recommendation algorithm not only considers the user’s behavior in a single domain,but also considers the user’s behavior in other domains and the interrelationship between these different domains.However,most of the existing cross-domain recommendation algorithms only use user rating data to obtain rating patterns to achieve recommendation,there is still a problem of rating data sparsity,which leads to inaccurate prediction of user preferences.In order to alleviate the sparsity of rating data,single-domain recommendation models often integrate review information to improve the recommendation effect.Inspired by this,this thesis designs two cross-domain recommendation methods that integrate rating and comment information.The main research work is as follows:(1)In view of the research on sparse data and cold start of users in most recommendation systems,this thesis designs a Cross-domain recommendation method based on CLFM model to deeply integrate scores and comments.Firstly,the extended stacked denoising autoencoder SDAE is used in the auxiliary domain and the target domain at the same time,and the rating information is used as the input of the SDAE model to realize the weighted fusion with the review information vector.Secondly,the orthogonal matrix tri-factorization technology is used to construct the rating patterns for the new rating matrix composed of the auxiliary domain,and the rating patterns in the auxiliary domain are transferred to the target domain with the help of the CBT transfer technology of the codebook.Finally,the rating model was adapted to the target domain and the final rating was predicted.The effectiveness of the proposed model was verified on several real datasets.(2)When multiple auxiliary information needs to be input at the same time,the SDAE model needs to increase the input and output layer dimensions.With the increase of the dimension of the output layer,it often leads to the loss of reconstructed data,thus affecting the prediction performance of the algorithm.By extending the cross-domain recommendation method of deep fusion of ratings and reviews,this thesis designs a Cross-domain recommendation algorithm based on semi-stacked denoising autoencoder.Because the semistacked noise reduction autoencoder SSDAE input layer and output layer dimension can be different.In this way,when the input dimension increases,the output dimension of SSDAE can remain unchanged,thus effectively alleviating the loss of reconstructed data.Firstly,the auxiliary and target domains use SSDAE models to complete the fusion of multiple input information.Secondly,for the potential feature vectors of users and items output by the middle layer,the extracted potential feature vectors are kept in the same space with the vector after matrix factorization by using the transformation function.Then,the orthogonal matrix tri-factorization technique is used to train the optimal migration matrix in the auxiliary domain,and the migration matrix is adapted to the target domain by combining the codebook-based migration technique.Finally,the final target domain rating matrix is reconstructed.In this thesis,we designed several comparative experiments on three real datasets,and the results show that the proposed model can improve the accuracy of recommendation to a certain extent.