Heterogeneous Information Network Embedding Based Cross-Domain Recommendation System

Along with the high-speed development of information technology,personalized recommendation system begins to play an extremely indispensable role in our daily life.Traditional recommendation systems require a large amount of interaction information between users and items to make the system more personalized.As a consequence,if the interaction information is inadequate,the performance of the system will be affected,which is known as the cold-start problem in personalized recommendation systems.Now days,there are two main research directions to solve the problem,heterogeneous information network(HIN)embedding based recommendation and crossdomain recommendation.The former one improves by introducing heterogeneous information to supplement the lack of interaction information.However,due to the insufficient data in the target domain,the improvement is limited.And also,the expression of users and projects based on the network structure cannot simultaneously extract personalized features.The latter one,cross-domain recommendation introduces auxiliary information from another domain,which has relatively dense dataset,and improve the effect of the system by transferring and integrating the auxiliary knowledge.But the type of the transferred knowledge is always single,so that the adaptability and scalability of the system are not good.In response to the above problems,we propose to combine heterogeneous information networks and cross-domain recommendation framework and to introduce heterogeneous information and cross-domain information to solve.Aiming at the two common recommend scenarios,score prediction and Top-K list recommendation,we propose HIN embedding based cross-domain frameworks respectively,and evaluate the effectiveness of them on real datasets.The specific contributions are as follows.(1)To solve the cold-start problem,this paper proposes to use heterogeneous and cross-domain information.We use tags as bridge to construct cross-domain heterogeneous information network,combine heterogeneous information network and cross-domain recommendations organically to achieve a better performance.(2)In the score prediction scenario,a fusion framework named Hec Rec is proposed to mine complex and diverse cross-domain heterogeneous information.Moreover,in order to avoid the knowledge conflicts conveyed by different meta-paths,we adopt the conception of "overpass" to process original embeddings.The experiments we conduct show that the absolute average error of Hec Rec is 0.6384,which is 2.7% reduced than the related work.(3)In the Top-K list recommendation scenario,we propose the fusion algorithm model EPCDRec to mine cross-domain heterogeneous information,which is based on embedding propagation layers and capable to study node embeddings characterizing both network structure features and personalized features in an end-to-end model.The precision,recall and cumulative normalized gain of the fusion model reach to 0.13,0.16 and 0.22 respectively,improved by 2.8%,1.2% and 3.0% compared with related work.14 figures,6 tables,and 52 reference articles are contained in the dissertation.