Sampling-based algorithm for link prediction in temporal networks

Abstract

The problem of link prediction in temporal networks has attracted considerable recent attention from various domains, such as sociology, anthropology, information science, and computer science. In this paper, we propose a fast similarity-based method to predict the potential links in temporal networks. In this method, we first combine the snapshots of the temporal network into a weighted graph. A proper damping factor is used to assign greater importance to more recent snapshots. Then, we construct a sub-graph centered at each node in the weighted graph by a random walk from the node. The sub-graph constructed consists of a set of paths starting from the given node. Because the similarity score is computed within such small sub-graphs centered at each node, the algorithm can greatly reduce the computation time. By choosing a proper number of sampled paths, we can restrict the error of the estimated similarities within a given threshold. While other random walk-based algorithms require O(n3) time for a network with n nodes, the computation time of our algorithm is O(n2), which is the lowest time complexity of a similarity-based link prediction algorithm. Moreover, because the proposed method integrates temporal and global topological information in the network, it can yield more accurate results. The experimental results on real networks show that our algorithm demonstrates the best or comparable quality results in less time than other methods.