Covert Networks: How to learn as much as possible about the structure of a network from sampled subnetworks

Covert networks of hostile actors, such as terrorists and cyberterrorists, constitute a serious national security concern. Effective counterterrorism strategies require understanding how covert networks form and operate. Statistical methods for social network analysis are powerful tools for shedding light on the structure of covert networks. An important and underresearched problem is that actors in covert networks may seek to hide connections and, while it may be feasible to observe some connections, it may not be possible to observe all of them. A related issue arises in many large networks: The size of large networks, combined with data collection constraints, may make it infeasible to observe all connections. That raises two related questions, which the PIs intend to address: I. Learning the structure of a network from small subnetworks: How should investigators learn the structure of a network from small subnetworks, while respecting the fact that subnetworks are embedded in the whole network of interest? II. Optimal sampling of subnetworks: Which subnetworks should an investigator seek to observe with a view to maximizing the information about the structure of the whole network? The PIs intend to address the two questions stated above as follows: Objective I: To estimate social network models from small subnetworks, the PIs will develop novel computational methods for maximum likelihood estimation that respect the fact that subnetworks are embedded in the whole network of interest. The PIs will focus on Exponential Random Graph Models (ERGMs), which belong to the richest and most popular classes of social network models. Compared with existing computational methods for maximum likelihood estimation of ERGMs, which exhibit too much variability to be useful for small subnetworks, the proposed computational methods have less variability and are expected to facilitate maximum likelihood estimation when more than 50% of the network data are missing. Objective II: The PIs will develop novel methods for sampling subnetworks with a view to maximizing the information about the structure of the whole network. Such methods are useful for: (a) Optimal sampling of subnetworks, where additional subnetworks are selected to maximize the information about the structure of the whole network. (b) Improving estimates of network effects, by using optimal sampling to observe more subnetworks. (c) Improved model-based predictions of unobserved subnetworks. Objective III: To demonstrate that the proposed methods are useful, the PIs will apply them to covert networks, with and without ground truth. Objective IV: The PIs will make all proposed methods available to social network analysts via R packages. The proposed methods will add new capabilities to the U.S. Army by providing sophisticated methods for optimal sampling from covert networks and learning the structure of such networks from sampled subnetworks. In addition, expert knowledge (e.g., human intelligence, signal intelligence) can be incorporated in the optimal sampling procedure. The resulting knowledge learned about the structure of such covert networks may help inform intervention strategies in covert networks, e.g., interventions aimed at dismantling terrorist networks.