Research Of Deep Cross-domain Recommendation Method Integrating Overlapping Users’ Review

Cross-domain recommendation algorithms achieve item recommendation for cold start users by migrating preference information from a source domain with rich user interaction records to a target domain with few or no user interaction records.In recent years,the crossdomain recommendation algorithm has developed rapidly and achieved good results in many application scenarios,but the algorithm is limited by the sparsity problem of the rating matrix,and most of the cross-domain recommendation algorithms adopt the interaction records of overlapping users in the source and target domains for cross-domain migration,but in practical application scenarios,the number of overlapping users is very sparse and most of the users’ comments are short,so it is difficult for the cross-domain recommendation algorithm to extract accurate user preference vectors from these short comments.In addition,the lack of interpretability of user preference feature vectors extracted by cross-domain recommendation algorithms is also a major difficulty faced by current recommendation algorithms.To this end,this thesis designs a cross-domain recommendation algorithm based on the information of overlapping user comments,and the research mainly includes the following two aspects:(1)A deep cross-domain recommendation algorithm incorporating overlapping user rating and comment information and multiple attention mechanisms is designed to address the long-standing data sparsity problem in cross-domain recommendation algorithms.First,based on the EMCDR algorithm,we add the review information of overlapping users in the source and target domains to improve the target domain rating prediction accuracy by using both user rating information and review information.Secondly,the comment text is processed at both local and global levels with the help of attention mechanism,so as to achieve the goal of focusing on the aspects that users really care about and ignoring the irrelevant aspects.Then convolutional neural networks are used for factor extraction to obtain factor vectors that more accurately represent user preferences.Finally,with the cross-domain migration method MLP of algorithm EMCDR,the migration of user preferences from the source domain to the target domain is realized.This algorithm is compared with the remaining six recommendation algorithms in six scenarios for experiments,and the experimental results prove that this algorithm effectively improves the recommendation effect.(2)The aspect-level deep cross-domain recommendation algorithm incorporating unoverlapping user comments is designed to address the problems of sparse number of overlapping users in the cross-domain recommendation scenario,short comments cannot accurately and comprehensively summarize user preferences,and the lack of interpretability of user preference vectors.Secondly,by using aspect-level gating mechanism and aspect-level attention mechanism,we assign weights to user feature vectors from aspect-level factor level as well as filtering to help improve the accuracy of user factor vectors and the interpretability of recommendation algorithms.Finally,the cross-domain migration method MLP is used to achieve cross-domain migration of user preferences by employing the algorithm EMCDR.This algorithm is compared with six recommendation algorithms in six scenarios,and the experimental results prove that the design of this algorithm effectively improves the recommendation effect.