Locating People of Interest in Social Networks

By representing relationships between social entities as a network, researchers can analyze them using a variety of powerful techniques. One key problem in social network analysis literature is identifying certain individuals (key players, most influential nodes) in a network. We consider the same problem in this dissertation, with the constraint that the individuals we are interested in identifying (People of Interest) are not necessarily the most important nodes in terms of the network structure. We propose an algorithm to find POIs, algorithms to collect data to find POIs, a framework to model POI behavior and an algorithm to predict POIs with guaranteed error rates. First, we propose a multi-objective optimization algorithm to find individuals who are expected to become stars in the future (rising stars), considering dynamic network data and multiple data types. Our algorithm outperforms the state of the art algorithm to find rising stars in academic data. Second, we propose two algorithms to collect data in a network crawling setting to locate POIs in dark networks. We consider potential errors that adversarial POIs can introduce to data collection process to hinder the analysis. We test and present our results on several real-world networks, and show that the proposed algorithms achieve up to a 340% improvement over the next best strategy. Next,We introduce the Adversarial Social Network Analysis game framework to model adversarial behavior of POIs towards a data collector in social networks. We run behavior experiments in Amazon Mechanical Turk and demonstrate the validity of the framework to study adversarial behavior by showing, 1) Participants understand their role, 2) Participants understand their objective in a game and, 3) Participants act as members of the adversarial group. Last, we show that node classification algorithms can be used to predict POIs in social networks. We then demonstrate how to utilize conformal prediction framework [103] to obtain guaranteed error bounds in POI prediction. Experimental results show that the Conformal Prediction framework can provide up to a 30% improvement in node classification algorithm accuracy while maintaining guaranteed error bounds on predictions