Research On Cross-domain Recommendation Based On Transfer Learning

With the popularization of internet technology and the arrival of big data era,people can obtain a lot of information through many channels,which brings people unlimited convenience while creating"information overload"problem,and the recommendation system can help people pick out valuable information from the complicated information,which can alleviate the"information overload"problem very well.However,traditional recommendation has problems of sparse data and cold start of users.Cross-domain recommendation can solve this problem very well.It transfers rich information from other fields to the target field and helps the target field improve the accuracy of the recommendation.However,most cross-domain recommendation work currently uses rating,ignoring review texts that contain richer user and project information.For this reason,this paper has carried out in-depth research on cross-domain recommendation methods,and the main work is as follows:?1?Considering that the recommendation system has the problems of sparse data and cold start of users,this paper proposes a cross-domain recommendation model based on convolutional neural networks?CNN-based Cross-domain Recommendation,CNCR?for review text.Firstly,the review text is divided into user review text document and item review text document,and corresponding word vector document are generated;Then use the Convolutional Neural Network?CNN?to effectively extract the rich information in the user and item review text documents to obtain user features and item features;Secondly,the idea of transfer learning is used to construct the shared domain as a bridge of knowledge transfer.The CNN of the shared domain is used to extract the shared features of the source and target domains,while the CNN of the source and target domains are used to extract the domain-specific features;Finally,in order to achieve cross-domain recommendation,the features of different fields are fused to make rating predictions.?2?Considering that the imbalance of the data will cause the model prediction results to be biased to the categories with a large number of samples data,this paper fuses the cost-sensitive learning method on the basis of the CNCR model,and proposes a CNN-based and cost-sensitive cross-domain recommendation model CNCR-CS?Cost-Sensitive and CNCR?.The CNCR-CS model uses a cost-sensitive learning method for the training process of the model,and proposes a L₂loss function CSL₂that incorporates a cost-sensitive strategy,which minimizes the model's misdivision cost and minimizes the above problems.?3?This paper conducted corresponding experiments on data sets in different fields of the Amazon to verify the effectiveness of the model.The experimental results show that,compared with the comparison model,the text model using the review text and neural network has improved in both MAE and RMSE,and achieved better rating prediction results.At the same time,the validity of the CSL₂function proposed in this paper and the effectiveness of the transfer method in this paper are verified.In addition,this paper designed other experiments by controlling the number of reviews to verify the model's ability to alleviate the data sparse problem and the user's cold start problem,and analyzed the important parameters of the model accordingly.At the same time,the generalization ability of the model is verified.This paper fully utilized the review text,reasonable and effective mining the review text contains user and project information,improve the performance of cross-domain recommended,at the same time makes cross-domain recommended method in deep learning and review text with new development.