Instance, Evolution, and Predictive Modeling of Social Networks

Abstract

Various phenomena within the information systems discipline can be studied using the social network paradigm that views social entities as nodes with links between them. The social network analysis (SNA) theory has applications in knowledge management, computer mediated communications, security informatics, and other domains. Challenges in SNA can be classified into three broad areas: instance modeling, evolution modeling, and predictive modeling. Instance modeling focuses on the study of static network properties, evolution modeling examines factors behind network growth, and predictive modeling is concerned with identification of hidden and future network links. This dissertation presents four essays that address these challenges with empirical studies in knowledge management and security informatics.The first essay on instance and evolution modeling contributes to SNA theory by examining a real-world network that contains interactions between thirty thousand individuals. The study is among the few that empirically examine large human-only networks and verify the presence of small-world properties and scale-free distributions. In addition, it proposes a novel application of a network evolution model to examine the growth of networks across geographical boundaries.The second essay on evolution modeling proposes a methodology to identify significant link-formation facilitators. The study found that homophily in age, gender, and race were not significant factors in predicting future links between individuals in dark networks. These results contradicted some previous studies in the same domain that used smaller datasets to study the phenomena.The third essay focuses on evolution and predictive modeling and examines the role of inventor status on the selection of knowledge recombined to produce innovation. A new network measure based on random walks and team identification (RWT) is proposed to model knowledge flow. It is found that inventor status as measured by RWT has a positive relationship with the likelihood of a future citation link to the inventor.The fourth essay focuses on predictive modeling. A modified mutual information formulation is proposed to identify hidden links between nodes based on heuristics of time and location of previous co-occurrences. An evaluation of the proposed technique showed that it performed better in predicting hidden links than other co-occurrence based methods.