Link Analysis In Social Networks

Link analysis is an important task in social network analysis. In mining social network, we always want more information about nodes, links and even communities.Link analysis is aiming to find semantic information about links. This paper mainly focus in the problems of link prediction and link classification.(1) Link prediction is a challenging task in social network analysis. In the problem of link prediction, we would like to find out the missing links and infer the future links in a snapshot of a social network. Previous studies focus more on the network itself sometimes with nodal side information. This paper presents a novel link prediction method, Personalized Pair-wise Link Matrix Factorization (PPLMF), which naturally combines hybrid factors including the social graph structure and the nodal personalized interests. We achieve this by formulating a probabilistic matrix factorization model incorporating with personalized interests information. A gradient based optimization method is used to solve the model. We conduct experiments on 2 real datasets, a co-authorship network and a direct social trust network. The experiment results show that our approach outperforms many state-of-the-art methods.(2) Link classification has not been discussed a lot. It is aiming to predict the type of social links like colleagues, friendship etc. Previous work has more focus in mining link information in ad-hoc domains. They did human feature engineering with experts’ knowledge. These algorithms are hard to be used in other domains. We prepose a feature representation algorithm combining with matrix factorization an Restricted Boltzmann Machine(RBM). It could extract hidden link features from social graphs. We conduct experiments in epinions and co-author network to prove our methods better than human feature engineering methods.Above all, out work in this paper is mainly focus in auto feature engineering, we use matrix factorization and RBM to extract hidden features from network and conduct experiments in real datasets to prove the efficiency of out methods.