ABSTRACT
Link prediction is one of the fundamental problems in network analysis. In many applications, notably in genetics, a partially observed network may not contain any negative examples, that is, edges known for certain to be absent, which creates a difficulty for existing supervised learning approaches. We develop a new method that treats the observed network as a sample of the true network with different sampling rates for positive (true edges) and negative (absent edges) examples. We obtain a relative ranking of potential links by their probabilities, using information on network topology as well as node covariates if available. The method relies on the intuitive assumption that if two pairs of nodes are similar, the probabilities of these pairs forming an edge are also similar. Empirically, the method performs well under many settings, including when the observed network is sparse. We apply the method to a protein–protein interaction network and a school friendship network.
KEYWORDS:
Acknowledgments
Yunpeng Zhao gratefully acknowledges the support from NSF DMS 1513004. Elizaveta Levina gratefully acknowledges the support from NSF DMS 01106772, NSF DMS 1159005, and NSF DMS 1521551. Ji Zhu gratefully acknowledges the support from NSF DMS 1407698, KLAS 130026507, and KLAS 130028612.