Link prediction in weighted networks via structural perturbations

Abstract

Link prediction aims at revealing missing and unknown information from observed network data, or predicting possible evolutions in near future. In recent years, extensive studies of link prediction algorithms have been performed on unweighted networks. However most empirical systems are necessarily to be described as weighted networks rather than solely the topology. In this paper we extend the structural perturbation method to weighted networks. We found that by including weight information the prediction accuracy can be significantly improved on networks with homogeneous weight distributions, meanwhile less improvements for heterogeneous weighted networks. Also we compared the weighted structural perturbation method to some benchmark algorithms, both weighted and unweighted, and found generally better performance in accuracy.