Research On Cross-Domain Recommendation Method Based On Feature Mapping

As a very effective solution to the information overload problem,recommender systems have been gaining more and more attention in recent years and are widely used in various applications,such as e-commerce systems and social networking systems,which greatly facilitate people’s lives.Although recommender systems have solved the information overload problem well,the cold-start problem has become a bottleneck to further improve their performance.In recent years,as the research on recommender systems deepens,how to solve the cold-start problem has gradually become one of the hot problems that major recommender system researchers pay much attention to,and a series of very effective models have emerged.As one of the ideas to alleviate the cold-start problem,cross-domain recommender systems have received more and more attention in recent years,and have achieved very good recommendation results.In this thesis,we focus on applying mapping-based cross-domain recommendation technology to solve the cold-start problem in recommendation systems,construct two mapping-based cross-domain recommendation models,and perform extensive experiments to verify the effectiveness of the models,and the research works are mainly as follows:A Cross-Domain Recommendation Model Based on Target Domain Feature Awareness and Complementary User Transfer.At present,mapping-based cross-domain recommendation technology has achieved very good results in solving the cold-start problem.The methods are mainly divided into two categories.One focuses on commonality,that is,all users share a mapping function;the other focuses on personality,that is,assign a personalized mapping function to each user;however,these two types of methods do not consider the complementarity of user commonality and personality;in addition,these two types of methods also ignore the mining of target domain knowledge itself.In order to solve the above problems,this thesis proposes a cross-domain recommendation model that combines complementary knowledge transfer and target domain feature extraction.On the one hand,a personality-commonality complementary mapping module is proposed to explicitly model the complementary information of user personality and commonality,and on the other hand,a relational network is proposed to mine target domain knowledge.Finally,the experiments on the Amazon dataset have achieved the best results so far,which verifies the effectiveness of the model proposed in this thesis.TJMN: Target-enhanced Joint Meta Network with Contrastive Learning for Crossdomain Recommendation.Cross-domain recommendation(CDR)provides a promising solution to mitigate the sparsity issue in the target domain by exploiting auxiliary information from the source domain.Recently,meta learning based methods have been proposed and achieved the state-of-the-art performance.However,these methods learn the transfer bridge solely relying on the source domain while the rich information from the target domain are ignored.Moreover,they leverage either a common transfer bridge or a personalized transfer bridge to transform users’ representations,without considering the multi-grained characteristics of user preference.In this thesis,we propose a target-enhanced joint meta network with contrastive learning(JTMN)for cross-domain recommendation.To be specific,we develop a target bridge to incorporate information from the target domain to guide the learning process of user preference transfer.In addition,we introduce multigrained transfer bridges to model the complex transfer patterns of user preference across different domains.At last,a target-aware contrastive learning layer is designed to obtain better user representations.The experimental results on three CDR tasks demonstrate that our proposed TJMN model significantly outperforms all strong baselines with a large margin,especially when the training data become more sparse.