Research On Multi-Relationship Aware Personalized Recommendation System

User-item interaction sequence information can reflect the evolution of users’ interests,more and more researchers are focusing on using this type of information to capture users’ interests for personalized recommendations.In real application scenarios,the items that the target user has interacted with are extremely limited,and it is difficult to dig out their true intentions only by using the user’s own interactive sequence information.Since the user’s social relationship information implies their potential interests,the use of the user’s explicit social relationship can enhance the target user’s expression,such as friend relationship,following relationship,and so on.In fact,there are implicit relationships in the social networks of users.Even if these users do not have an explicit social relationship,they may also have common interests.However,the implicit relationships between users are diverse and complex,such as fans of a certain blogger,comment users of a certain Weibo,users who make similar comments,etc.The types of implicit relationships between them are different and the degree of mutual influence is different.How to accurately express and use such relationships is still a challenging research topic.A novel personalized recommendation model based on multi-relationship discovery is proposed in this paper,starting with the construction and mining of the relationships between users,it aggregates multiple relationships among social users from the aspects of user personal information and user interaction sequence.Specifically,the following solutions are adopted:(1)Construct user social relationship diagrams,user interaction diagrams,user potential relationship diagrams based on personal information,and user potential relationship diagrams based on interaction sequences,according to the heterogeneous diagrams of user-item,to more comprehensively express the relationship between social users Multiple relationships.(2)Based on the above four relational graphs,design an aggregation method based on graph neural network to capture the mutual influence of interest between users,and use an attention network to adaptively distinguish between different relations The degree of importance is used to improve the reliability of aggregation of multiple relationships.On the other hand,the weight of personal information and interaction sequence is adaptively controlled to determine the user’s final expression under the influence of multiple relationships.This paper innovatively constructs a variety of relationship graphs based on social network information,users’ personal information and interaction sequences,and proposes a representation learning method based on graph neural networks and double-layer attention networks.Compared with existing methods,it provides a more comprehensive expression and The display relationship and implicit relationship between users are used to alleviate the sparse problem of single user interaction items,better characterize the user’s interest characteristics,and be compatible with existing sequence recommendation methods.Experiments on two public data sets show that the performance indicators of HR@10 and NDCG@10 are better than the existing methods,verifying the superiority of the model.