Inferring Unobservable Inter-community Links in Large Social Networks

Abstract

Social networks can be used to model social interactions between individuals. In many circumstances, not all interactions between individuals are observed. In such cases, a social network is constructed with the data that has been observed, as this is the best one can do. Recent research has attempted to predict future links in a social network, though this has proven a very challenging task. Rather than predicting future links, we propose an inference method for recovering the links in a social network that already exist but that have not been observed. In addition, our approach automatically identifies groups of individuals that form tight-knit communities and models the intra and inter-community interactions. At this higher level of abstraction and for a social network built frommobile phone calls, our method is able to accurately identify a subset of 10% of all community pairs where about 50% of the pairs have had unobserved communication between them, an improvement of about four times over a subset of the same size with randomly chosen pairs. To the best of our knowledge, this is the first method that infers links that exist but are unobservable in a phone call-based social network. In addition, we perform the inference at the community level, where the discovery of unobserved inter-community communication can provide further insight into the organizational structure of the social network andcan identify social groups that may share common interests.