Identifying Covert Sub-Networks Through Iterative Node Classification

In this paper we study a problem of identifying a group of related individuals embedded in a larger population. We state the problem in terms of node classification in a social network, and present an iterative algorithm to classify individuals. We test it empirically on data generated by the Hats simulator. Despite its simplicity, the algorithm performs remarkably well. Like most iterative processes, iterative classification has characteristic dynamics. We demonstrate that the dynamics of classifying group members differs from the dynamics of classifying non-members. We call this phenomenon two-tiered dynamics. Our algorithm exploits this difference to identify group members with high accuracy.