| In the Internet age,the massiveness of data has caused information overload.Recommender systems solve the problem of information overload by providing users with useful item suggestions.However,in real-world recommender systems,data sparsity has always been a challenging and common problem,which can affect the accuracy of recommendations and,in turn,user experience.Cross-domain recommender systems aim to leverage knowledge in auxiliary domains with relatively abundant data to improve recommendation performance in target domain with sparse data.Transfer learning is a key technique to achieve cross-domain recommendation.It can transfer useful knowledge in the auxiliary domain to the target domain to solve the problem of data sparsity in the target domain.This dissertation mainly focuses on the problems of insufficient knowledge transfer,negative transfer and inconsistent knowledge transfer in knowledge transfer-driven crossdomain recommender systems under the scenario of extremely sparse data.The main innovations are as follows:(1)For the scenario where users/items between domains completely overlap and the rating forms are heterogeneous,we propose a cross-domain recommendation algorithm,named Enhanced Knowledge Transfer for Collaborative Filtering with Heterogeneous Feedbacks(EKT),to address the problems of insufficient knowledge transfer and negative transfer.On the basis of the latent factor transfer model,a partial cluster-level rating pattern transfer is constructed to enhance the knowledge transfer ability of the algorithm,and the graph co-regularization of the user graphs and the item graphs is used to maintain the inherent geometric structure of the domain and avoid negative transfer.Finally,an efficient optimization iterative algorithm is proposed and its convergence is proved.Experimental results on two real-world datasets show that the algorithm outperforms the comparison methods when the density is low than 1%.(2)For the scenario where only items overlap between domains,we propose a crossdomain recommendation algorithm,named Selective Knowledge Transfer for Cross-domain Collaborative Recommendation(SKT),to solve the problems of insufficient knowledge transfer and negative transfer.On the basis of the item latent factor transfer model,the selective transfer of user latent factors is constructed to avoid negative transfer while enhancing knowledge transfer,and the graph co-regularization of user and item graphs is used to further avoid negative transfer.Experimental results on two real-world datasets show that the algorithm outperforms the comparison methods when the density is low than or equal to1%.(3)For the scenario of non-overlapping entities between domains,we propose a crossdomain recommendation algorithm for the problem of inconsistent knowledge transfer and negative transfer,named Cross-domain Collaborative Recommendation without Overlapping Entities Based on Domain Adaptation(CCR-DA).Based on the cluster-level rating patternbased transfer model,the constructed maximum mean discrepancy is used to align the feature distributions of users/items across domains to achieve consistent knowledge transfer,while graph co-regularization is used to avoid negative transfer.Experimental results on three types of tasks show that the performance of the algorithm is better than that of the comparison methods,especially when the density is lower,the performance improvement is greater.(4)For the scenario of no overlapping entities between domains,we propose a crossdomain recommendation algorithm for the problems of insufficient knowledge transfer,inconsistent knowledge transfer and negative transfer,named Cross-Domain Recommendation with Multi-auxiliary Sources via Consistent and Selective Knowledge Transfer(MS-CSKT).On the basis of the cluster-level knowledge transfer model based on multiple auxiliary domains,the constructed maximum mean discrepancy regularization term is used to align the feature distribution of users/items in multiple auxiliary domains and the target domain,so as to enhance the knowledge transfer and avoid the inconsistency of knowledge transfer.Simultaneously construct selective cluster-level knowledge transfer to avoid negative transfer.Furthermore,graph co-regularization of user and item graphs is further used to avoid negative transfer.Experimental results show that the recommendation performance of the algorithm is better than that of the comparison methods,and the performance improvement is more significant when the density is lower. |
