Link prediction for existing links in dynamic networks based on the attraction force

Abstract

Link prediction has become an important technology which can be widely applied to many aspects of social network analysis. Numerous link prediction methods have been proposed in recent years, which can detect missing links in the current network and predict whether two disconnected nodes will be connected in the future. However, some existing links may be spurious links that should be removed from the current network, and some existing links will leave the network in the future because network changes over time. In this paper, a novel link prediction approach based on the attraction force between nodes in dynamic networks is proposed to detect spurious links at time t and predict whether existing links at time t will leave the network at time t + 1. At first, the attraction force between nodes is defined according to the law of gravitation, and a level is assigned to each node in the network at time t. Then, the connection probability of each pair of connected nodes is calculated according to the level of nodes and the attraction force between nodes. Next, the connection probability of all existing links is calculated to detect whether they were spurious links that should be removed from the network at time t. For an existing link, the lower the connection probability, the more likely it is to be a spurious link. Furthermore, the changes of network from time t to t + 1 is considered to build a virtual network and the connection probability of all existing links in the virtual network is calculated to predict whether they will leave the network at time t + 1. Experimental results on some real-world networks show that the proposed approach can achieve acceptable prediction accuracy and can be effectively applied to the prediction of existing links in dynamic networks.