User Link Analysis Based On Multiple Social Networks

With the rapid development of the information age,internet has became an important part of people's life.Meanwhile,the online social network even been more popular with an explosive trend,people tend to share their daily life,entertainment and work et al in the online social network.Social links,as the ties which connect users,are the fundamental part of the social network.Analysis on the links of online social network has been attracted more and more researchers' interests.Analysis on the links of online social network can benefit amount of social network applications such as link recommendation,online advertisement,information diffusion and public sentiment monitor.At present,most studies on the social link analysis are based on the individual network.In fact,there are several multiple types of social networks and users are taking part in different social networks.Analysis on links of multiple social network not only can deeply understand the characteristic of links increment,provide some useful suggestions for network developers and user for attracting and maintain social links,but also can alleviate the sparsity issue of link prediction and "cold start" problem in recommendation.But there several challenges in analysis on links of multiple social network: due to the different properties between social networks,the reasons for the social link formation are different.For example,users in microblog system from the link with others for obtaining self-interesting information,but users in academic social network form the links with others because they have similar research fields or they are in the same institute.Besides,it's hard to identified the same user in multiple social networks due to the closure between them.Focus on the problems above,based on the link of multiple social networks,we define two types of links: social links in individual network and anchor links between multiple social networks.We try to model these two types of links using the content and structural information.The contribution of our works are summarized as follows:Analysis on the speed of social links increment of multiple social networks.There exist different properties in multiple social networks.Which content and structural features users have in multiple social networks can induce a fast speed of attracting new social links is the primary goal of our study.In this thesis,based on users' content and structural information,we define the variables such as diversity and density,then we analyze the correlations between users' speed of attracting social links and these variables.The experimental results which are conducted on two real world social network-Weibo and Aminer show that: due to different properties of multiple social network,there are some differences between users in multiple social network for attracting new social links.These analysis can provide some useful suggestions for network developers and users for attracting and maintain social links.Ranking people who get social links fast in a short time: Attracting friends(or followers)in a short time is a strong indicator of one person for becoming an influential user quickly.Finding users who get social links fast can provide efficient and economic seed users for social analysis applications such as influence maximization.We proposed a partially-labeled ranking factor graph model(PLR-FGM)in this thesis.Two kinds of factor functions are defined in the model.The attribute factors are used to represent the properties of users' contents and structures.The correlation factors are defined to capture the users' relationship in the network.The experimental results on two different social networks show that,different features have different contributions for the ranking model.Besides,the proposed model in this thesis outperforms several state-of-the-art models which can not model the users' relationships.Anchor link prediction in multiple social networks: mapping users in multiple social networks can alleviate the sparsity issue and transfer useful information and benefit applications like social link prediction and cross-domain recommendation.In the anchor link prediction,it is hard to get the ground-truth anchor links,the methods which are based on supervised classification will encounter with the imbalance problems.Moreover,most of the existing work carry out structural alignment of networks with matrix factorization and inverse involved,making them hard to scale up for large-scale problems.A representation learning model with the objective to learn an aligned network embedding for multiple networks is proposed.Stochastic gradient descent and negative sampling are used for the efficient learning of the model.To map users across networks,we compute the cosine similarity between the vector representations of one node in different networks to determine the correspondence.