Research On Social Network Structure Features Based Social Recommendation System

With the rapid development of mobile Internet,the amount of information available on the Internet has caused a serious information overload problem for users.Internet Users are unable to distinguish what is important.The recommendation system has effectively solved this problem.The traditional recommendation system is limited by the inherent problems of sparse user behavior data and cold start,therefore it cannot provide accurate and reasonable recommendation results.With the continuous development of social networks,social data has become an important available data source,and social recommendation systems dedicated to mining users’ social data to improve the performance of recommendation systems have emerged.In this thesis,we study how to use social information data from online social networks to improve the recommendation performance of recommendation systems,specifically,by learning the structural feature information of social networks.The specific research contents of this thesis are as follows:(1)To address the problems that traditional social recommendation system models have difficulty in mining indirect,higher-order dependencies in social networks,cannot process multiple social information simultaneously,and the fusion of social information cannot provide personalized recommendation results,this thesis proposes a recommendation model based on Node2 Vec global social network embedding,and a latent factor model for rating prediction.The model combines the latent factor model,the Node2 Vec network embedding representation technique,and the two different kinds of fusion approach of social information.Node2 Vec,as a general network embedding representation learning technique,enables the model to fully explore local and global social information while being able to handle different types of social information,and the two different kinds of fusion approach of social information ensures that the two different types of fusion approaches of social information ensures that the recommendation results provided by the model can meet the personalized needs of users.(2)To address the problem that the improvement of the social embedding representation vector directly fused into the matrix decomposition model enhancement is very limited,this thesis further proposes the use of a meta-path based random walk approach,combined with the embedding part of the Node2 Vec method,to embed different types of nodes in the heterogeneous network representation,so as to reach the filtering of users’ social information at the interest level,to make the recommendation performance Further improvement.In terms of fusion methods for social information,a fusing function of social information based on attention mechanism is proposed.Finally,a large number of comparative experiments are conducted on real datasets,and the results show that both proposed methods,compared with previous models of social recommendation systems,achieve better recommendation performance and effectively alleviate the data sparsity and cold start problems of recommendation systems,fully illustrating the rationality and effectiveness of the proposed algorithms.