Research On Cross Domain Recommendation Methods Based On Deep Learning

With the wide spread and rapid development of the Internet,recommendation methods have been applied to many recommendation scenarios,enhancing user experience and also bringing new opportunities for research in cross-domain recommendation methods.Crossdomain recommendation methods that transfer knowledge from the source domain to the target domain to enrich the latter’s information perform better than single-domain recommendation methods.In cross-domain recommendation methods,the transfer and fusion of heterogeneous auxiliary domain user-item features is crucial,and also an important issue that affects rating prediction and item recommendation.Therefore,this paper studied feature extraction of useritem characteristics in features’ migration and fusion of cross-domain algorithms of single domain.The main contributions of this paper is as follows:Firstly,to address the problem of data sparsity in the feature extraction algorithm of deep mixed collaborative filtering in a single domain,we proposed a neighborhood semi-matrix factorization-based algorithm(NSe MF)that integrated neighbor preference auxiliary information and a Semi-Auto Encoder to effectively extract user and item features.This method filled in the missing values of the original rating matrix in both horizontal and vertical dimensions through the MC-WKNN method,which introduced neighbor preference auxiliary information,and then combined a Semi-Auto Encoder to introduce item side auxiliary information with Matrix Factorization for ultimately extracting effective features of users and items.Experimental results showed that this method can reduce error and improves recommendation performance.Secondly,to address the negative transfer problem that arises when integrating multiple heterogeneous auxiliary domain features in cross-domain recommendation methods,we proposed a multi-head attention DeepFM cross-domain recommendation(MADF-CDR)method based on a multi-head attention mechanism.This method aimed to automatically assign optimal weights to the auxiliary domains during feature transfer and fusion to accurately predict target domain ratings.The model first used the NSe MF method to extract user and item features from multiple domains,and then utilized the multi-head attention mechanism and the crossproject domain idea to improve the DeepFM algorithm for feature transfer and fusion.The FM module and DNN module were used to learn the first-order,second-order and high-order nonlinear interactions of features to obtain the target domain’s predicted ratings and recommendation results.This model fully utilized the effective features of multiple auxiliary domains and significantly improved cross-domain recommendation performance.Finally,comparative experimental results show that the NSe MF algorithm,which integrated neighbor preference auxiliary information,has lower error compared to existing methods,and the multi-attention DeepFM cross-domain recommendation model has higher prediction accuracy and recall results.