A Supervised Learning Approach To Link Prediction In Dynamic Networks

Link prediction, as one of fundamental problems in social network, has aroused the vast majority of research on it. However, most of existing methods have focused on the static networks, although there exist some machine learning methods for the dynamic networks, they regard either link structures or node attributes captured from a single snapshot of the network as the features, thus cannot achieve high accuracy. In this paper, following the supervised learning framework, we innovatively propose a new approach to this problem in dynamic networks. In particular, our features are captured from the variation of the structural properties and a lot of important metrics considering the longterm graph evolution of network, instead of a single snapshot. For each feature, we use an optimization algorithm to calculate the corresponding weight of each classifier, and then can determine whether there is a connection between a pair of nodes. In addition, we execute our method on two real-world dynamic networks, which indicate that our method works well and significantly outperforms the prior methods.