Research On Personalized Recommendation Algorithm Based On Graph Convolution Network

The rapid development of the Internet can help people out of the once closed environment of information,so that we can know the world more conveniently.However,the information on the network is increasingly mixed.It is a test for people to manually find their real needs in such complicated information.The recommendation system is proposed to solve this problem.In the development process of recommendation system,based on the powerful characterization ability of graph convolutional network,the combination of recommendation system and graph convolutional network has gradually increased.The addition of graph convolutional network can help the recommendation system dig deeper into the potential value of users and items,and effectively alleviate the problems of data sparsity and cold start.After studying the research methods related to recommendation system and graph convolution network,this paper proposes a multi-task recommendation algorithm based on Deep FM and graph convolutional network(Deep FM_GCN).In addition,the characteristics of users and items are not immutable and will change with time.Therefore,in order to understand the relevant research of some dynamic recommendation systems,this paper proposes a double recurrent recommendation algorithm based on the dynamic lightweight graph convolutional network(DR-DLGCN).The specific research contents of this paper are as follows:(1)Aiming at the problem that the sparsity of user-item data leads to poor recommendation effect,an item-centered knowledge graph is constructed as additional supplementary information of the item,and multi-task alternative learning is conducted with the recommendation task.(2)An information sharing module is designed to connect the entity information of the knowledge graph with the item information entered in the recommendation task,and supplement the information of the other party.In the knowledge graph representation learning,the graph convolution network is used to model the knowledge graph,making full use of the entity information and relationship information in the graph to avoid semantic deviation caused by insufficient node representation learning.(3)For most sequence recommendations,only a single user’s historical interaction behavior sequence is modeled,ignoring different user’s historical interaction behavior sequences and the existence of higher-order connectivity information between each item in the behavior sequence,the user-item dynamic graph is constructed,and the graph convolution network is used to extract this higher-order connectivity information.And considering the complexity of the graph convolution network,the lightweight graph convolution network is introduced to mine the sister graph node information.(4)In view of the problem that most of the sequence recommendations only consider the dynamic preferences of users,this paper constructs the scoring behavior sequence of users and items to learn the dynamic characteristics of users and items.The user sequence can update the user’s dynamic preferences according to the interactive item features in the sequence,and the item sequence can also update the item’s dynamic features according to the interactive user features in the sequence.At the same time,consider the score value between users and items as the weight to measure the importance of each node information in the sequence.