Research On User Relationship Discovery Of Signed Network Based On Nonnegative Matrix Factorization

With the rapid development of social networks,obtaining information between users from social networks has become a research hotspot in the academic field,such as link prediction,community detection and node classification.Social networks often have large number of users,and the links between users are extremely sparse.In traditional graph theory,the adjacency matrix is used to represent social network.And the adjacency matrix is very large,and most of the data are zero,so it is difficult to effectively use social networks.This thesis proposes an Orthogonal Graph regularized Nonnegative Matrix Factorization(OGNMF)for signed networks,an important branch of social networks.The OGNMF can reduce the dimension of signed network,make its dimension suitable for storage and calculation and retain the structure information and sparsity at the same time,the processed signed network can be used to predict the relationship between users.In addition,we use 2,1 norm on the basis of OGNMF as constraint to make it more robust.The main tasks of this thesis are described as follow:1.We propose the OGNMF algorithm and provide its updating rules.With these rules,the OGNMF algorithm can achieve convergence.In addition,we also provide the analysis of convergence.The OGNMF algorithm can effectively reduce the dimension of high-dimensional social network,so that the low-dimensional matrix can be used to replace the original matrix to compute.User relationship in signed network is regarded as label,in scenarios where labels are insufficient,firstly we reduce the dimensions of the source domain network and the target domain network at the same time,then we use the Tr Ada Boost,an algorithm in transfer learning,to train the samples in the source domain network and the target domain network together.The Tr Ada Boost can increase the weight of useful samples and reduce the weight of useless samples automatically.The samples after training can be used in data mining.We have carried out multiple experiments on three open datasets,Wiki-RFA,Slashdot and Epinions.Compared with four groups compared algorithms,our algorithm can achieve the best results.2.To solve the problem of class label noise in signed networks,this thesis proposes a Robust Orthogonal Graph regularized Nonnegative Matrix Factorization(ROGNMF)method based on OGNMF.The ROGNMF uses 2,1 norm as constraints in the objective function of OGNMF to reduce the interference of outliers or noise points on the original data distribution.We also give the update rules and analysis.According to the different noise rates,we have carried out many experiments on three datasets.The experimental results show that ROGNMF algorithm can effectively reduce the impact of noise points.