New approaches to maximizing influence in large-scale social networks

Abstract

With the widespread adoption of social media in today's society, the problem of identifying the most influential individuals whose adoption of a product or action will spread maximally in the network is of increased practical significance. This thesis considers new strategies and methods for this problem, which is known as the influence maximization problem, focusing on a setting where the influence is determined by some function of user opinions. In the first chapter, we introduce a new model of opinion dynamics that is motivated by research in both social psychology and political science. We present a series of theoretical results concerning the convergence of the opinions to an equilibrium, including conditions under which convergence to a fixed point occurs, an explicit characterization of the equilibrium, and the rate of convergence to the equilibrium. In the second chapter, we propose new approaches to the influence maximization problem in a social network when the dynamics adhere to the model in the first chapter. We consider applying these methods on several large-scale real-world social networks. In doing so, we attempt to measure the validity of the model we propose, consider estimating the relative importance of some special users via a centrality function approach, and highlight the computational efficiency of our influence maximization methods. In the final chapter, we introduce an alternative approach to maximizing influence in a social network that has as a solution a dynamic policy that considers when, what, and with whom an agent communicates. We motivate the necessity for a dynamic policy solution by highlighting some realistic behaviors that make modeling and analyzing real-world dynamics difficult. By leveraging reinforcement learning, we learn policies that account for some of these realistic behaviors and find that these policies exhibit impressive performance on large-scale networks.