Research On Transfer Learning In E-commerce Cross-domain Recommendation

The rapid development of the Internet and big data technology has brought people the convenience of learning and life.At the same time,it has also caused people to suffer from information overload problem.People cannot quickly obtain interested information from vast amounts of online data.In this context,personalized recommendation technology came into being,which can solve the problem of information overload to a certain degree.However,currently widely used recommendation technologies are concentrated in a certain area,which may lead to data sparsity and cold start issues.Researches and facts show that users show the same preferences in different areas.For example,users who like inferential movies also tend to purchase reasoning books.This dissertation refers to the field containing rich user and rating information as the source domain,and the field of data sparseness as the target domain.How to use the information in source domain to provide more accurate recommendations for users in target domain is the main research content of this dissertation.Starting from the research of personalized recommendation,cross-domain recommendation and transfer learning,this dissertation focuses on the problems of data sparsity and cold start in single domain recommendation.Two algorithms are proposed in this dissertation,which one based on tag migration learning and another based on user interest.In the first algorithm,shared label in two domains are used as a bridge of two domains,and for the source domain,non-negative matrix factorization and adopted K-means algorithms are used in user-item matrix.Processes and migrates the tag shared in two domains and presents the user classification results in the target domain,and finally the same type of users is used to populate the unrated items in the target domain with then mean of ratings.However,it's difficult for different domains sharing the same labels.So,a cross-domain recommendation algorithm based on user interest is proposed in this dissertation when two domains have some same users.The source domain user interest model and the target domain user interest model are established using singular value decomposition and adopted similarity algorithm.And then,the target domain user prediction score is given to provide product recommendation for the target domain user.This dissertation uses the public data set MovieLens and Alibaba Cloud Tianchi database set to experiment respectively.The experimental results show that the cross-domain recommendation algorithm based on tag migration learning can solve the problem of data sparsity and cold start to a certain extent,and it is beneficial to solve inter-domain problems,and that cross-domain recommendation algorithm based on user interest also alleviates data sparsity and cold start problems and can provide users with more accurate recommendation results.