Node Classification In Online Social Networks Based On Manifold Learning

Along with the rapid development of Internet and coming of Web 2.0 age, online social network user scale is increasing. Large amounts of networked data are produced. Mining network hidden rules and patterns from the complicated structure information by using data mining and machine learning technology, has very important theoretical significance and broad application prospect in understanding the function of network, the latent connection and behavior between users.Node classification is a very important problem in social network data mining. In node classification tasks, parts of the nodes have labels that reflected users’ interest and behavior.The goal is to label the others by using structure information. Latent social dimension model is a good way to solve the problem of heterogeneous relationship between nodes, and uses soft clustering algorithm modularity maximization for social dimension extraction. But its classification accuracy is quite low. Laplacian Eigenmaps is a typical method in manifold learning.It has an implicit emphasis on clustering characteristics of the data and can be interpreted as a soft clustering algorithm. This paper proposes a node classification algorithm which uses Laplacian Eigenmaps to extract social dimensions. Experiments show that the classification performance is better than that of the original social dimension model. This paper also proposes an improved algorithm which combined the content information and the classification accuracy can be further improved. The algorithm proposed by this paper can capture implicit user relations better so as to analyze users’ interest and behavior better.In addition, data acquisition from the Web is an effective data source to verify node classification algorithm and is one of the important function links for developing an online network data analysis system. Therefore, this paper has implemented data acquisition from two typical social network sites Netease and Renren on the basis of analyzing their network structure. It has been applied to the system developed by our research group.