SPINN: Suspicion prediction in nuclear networks

The best known analyses to date of nuclear proliferation networks are qualitative analyses of networks consisting of just hundreds of nodes and edges. We propose SPINN - a computational framework that performs the following tasks. Starting from existing lists of sanctioned entities, SPINN automatically builds a highly augmented network by scraping connections between individuals, companies, and government organizations from sources like LinkedIN and public company data from Bloomberg. By analyzing this open source information alone, we have built up a network of over 74K nodes and 1.09M edges, containing a smaller whitelist and a blacklist. We develop numerous features' of nodes in such networks that take both intrinsic node properties and network properties into account, and based on these, we develop methods to classify previously unclassified nodes as suspicious or unsuspicious. On 10-fold cross validation on ground truth data, we obtain a Matthews Correlation Coefficient for our best classifier of just over 0.9. We show that of the 10 most relevant features for distinguishing between suspicious and non-suspicious nodes, the top 8 are network related measures including a novel notion of suspicion rank.