TeleLink: Link Prediction in Social Network Based on Multiplex Cohesive Structures

Abstract

Given a network where the same set of nodes have multiple types of relationships, how do we efficiently predict potential links in the future (e.g., interactions between social actors), and how do we predict links using information from other relationships? These problems have been widely studied recently, most of the existing methods either aggregate multiple types of relationships into a single network or consider them separately and ignore the correlations across relationships, leading to information loss. In this work, we present TeleLink, a general link prediction model that works for networks with single and multiple relationships. TeleLink predicts potential links based on community detection and improves link prediction by bringing in a cohesive structure across multiple networks constructed by different relationships or node attributes. To further improve the prediction performance, we extend TeleLink to a semi-supervised scheme, incorporating partially labeled information. Our extensive experiments show that TeleLink outperforms existing methods in predicting new links. Specifically, among the various datasets that we study, TeleLink achieves a precision improvement by up to 110 % compared to the baselines.