Date of Award

2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Margaret J. Eppstein

Abstract

Network science captures a broad range of problems related to things (nodes) and relationships between them (edges). This dissertation explores real-world network problems in disparate domain applications where exploring less obvious "hidden networks" reveals important dynamics of the original network.

The power grid is an explicit network of buses (e.g., generators) connected by branches (e.g., transmission lines). In rare cases, if k branches (a k-set) fail simultaneously, a cascading blackout may ensue; we refer to such k-sets as "defective". We calculate system risk of cascading failure due to defective 2-sets and 3-sets in synthetic test cases of the Polish and Western US power grids. A stochastic group testing algorithm (Random Chemistry) is used to efficiently sample defective k-sets in the "hidden network" of all possible k-sets, and new methods are proposed to derive bounds on the total number of defective sets from the obtained sample. We use copula analysis, with a custom distance metric, to estimate risk when the k initiating outages are spatially correlated and show that this realistic assumption increases the relative contribution to risk of 3-sets over 2-sets.

In the power systems application, among others, computational costs vary when testing defective vs. non-defective k-sets, a consideration that has not previously been made when evaluating group testing algorithms. We develop a domain-independent test problem generator that enables us to vary the number of defective k-sets, with a tunable parameter to control the cost ratio of testing defective vs. non-defective k-sets. We introduce a deterministic group-testing algorithm (SIGHT) capable of sampling from this space, and show that both the number of defective sets and the test cost ratio affect the relative efficiency of Random Chemistry vs. SIGHT. We discuss various applications where each algorithm is expected to outperform the other.

Conversations can also be viewed as explicit networks of dialog (edges) between speakers (nodes). We propose using second and third order Markov models based on the sequence of speaker turn lengths to elucidate "hidden networks" of information flow and reveal patterns of information sharing between participants. The proposed method is demonstrated on a corpus of conversations between patients with advanced cancer and palliative care clinicians. We demonstrate the efficacy of the model by confirming known patterns of conversational discourse, identifying normative patterns of information flow in serious illness conversation, and showing how these patterns differ under a variety of contexts, including the expression of distressing emotion.

Language

en

Number of Pages

217 p.

Share

COinS