Diffusion Dynamics in Interconnected Communities

Abstract

In this dissertation, multi-community-based Susceptible-Infected-Recovered (SIR) and Susceptible-Infected-Susceptible (SIS) models of infection/innovation diffusion are introduced for heterogeneous social networks in which agents are viewed as belonging to one of a finite number of communities. Agents are assumed to have well-mixed interactions within and between communities. The communities are connected through a backbone graph which defines an overall network structure for the models. The models are used to determine conditions for outbreak of an initial infection. The role of the strengths of the connections between communities in the development of an outbreak as well as long-term behavior of the diffusion is also studied. Percolation theory is brought to bear on these questions as an independent approach separate from the main dynamic multi-community modeling approach of the dissertation. Results obtained using both approaches are compared and found to be in agreement in the limit of infinitely large populations in all communities. Based on the proposed models, three classes of marketing problems are formulated and studied: referral marketing, seeding marketing and dynamic marketing. It is found that referral marketing can be optimized relatively easily because the associated optimization problem can be formulated as a convex optimization. Also, both seeding marketing and dynamic marketing are shown to enjoy a useful property, namely ``continuous monotone submodularity." Based on this property, a greedy heuristic is proposed which yields solutions with approximation ratio no less than 1-1/e. Also, dynamic marketing for SIS models is reformulated into an equivalent convex optimization to obtain an optimal solution. Both cost minimization and trade-off of cost and profit are analyzed. Next, the proposed modeling framework is applied to study competition of multiple companies in marketing of similar products. Marketing of two classes of such products are considered, namely marketing of durable consumer goods (DCG) and fast-moving consumer goods (FMCG). It is shown that an epsilon-equilibrium exists in the DCG marketing game and a pure Nash equilibrium exists in the FMCG marketing game. The Price of Anarchy (PoA) in both marketing games is found to be bounded by 2. Also, it is shown that any two Nash equilibria for the FMCG marketing game agree almost everywhere, and a distributed algorithm converging to the Nash equilibrium is designed for the FMCG marketing game. Finally, a preliminary investigation is carried out to explore possible concepts of network centrality for diffusions. In a diffusion process, the centrality of a node should reflect the influence that the node has on the network over time. Among the preliminary observations in this work, it is found that when an infection does not break out, diffusion centrality is closely related to Katz centrality; when an infection does break out, diffusion centrality is closely related to eigenvector centrality.