Network Embedding And Its Application In Recommender System

With social networks and e-commerce platforms becoming an indispensable part of life,how to provide users with interested recommendation items has become a key issue for major platforms to study.For the user-item bipartite graph of the recommender system,it is necessary to mine low dimensional vectors that can accurately represent users and items,so the research of Web-based representation learning method is increasingly active.The existing network embedding method and its application in recommender systems still have the following problems: 1.Only focusing on the local structure of networks,ignoring the global topology,we need to add symmetric constraint to optimize node representation.2.The performance of recommender system needs to be improved,and collaborative filtering based on single element matrix factorization needs to combine node attributes and behavioral history to improve prediction accuracy.Traditional hybrid recommendation needs to combine the advantages of single element matrix factorization to reduce time consumption.In view of the above problems,two algorithm models are proposed by this thesis to improve the learning performance of network representation:1.In order to preserve the network topology better,we propose a learning method of network representation based on the alternating direction multiplier(ADMM),symmetric and non-negative constraints.First,we propose the objective function with symmetry restriction,the network representation of nodes is constrained by modularity,community attribute,symmetry and similarity between nodes.Next,we deduce the updating formula of unknown elements in the objective function by using the alternating direction multiplier method of single element.Finally,the clustering and visualization of the updated network representation results are carried out,and the results on the real dataset prove the validity of our model.2.In order to solve the problems of prediction accuracy and time consumption in recommendation scenarios,we propose a learning method of network representation based on non-negative matrix factorization(NMF),which is suitable for hybrid recommender systems.Firstly,we calculate the user similarity matrix and item similarity matrix by user attributes,behavior history and item attributes.Then,the objective function is derived by single element based non-negative matrix factorization method,and the representation matrices of user and item are obtained.Finally,the similarity between user representation and item representation is calculated to recommend the interested itemts for users.3.Based on the proposed hybrid recommender system algorithm,a recommender system framework is designed and implemented for movie recommendation scene.Specifically,it includes data analysis and preprocessing,vectorization of various types of data,combined with the web development framework of Django + Vue + bootstrap,and finally completes accurate movie recommendation for users.