Date of Award

Spring 2026

Language

English

Embargo Period

5-15-2027

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College/School/Department

Department of Computer Science

Program

Computer Science

First Advisor

Dr. Abram Magner

Committee Members

Dr. Petko Bogdanov, Dr. Charalampos Chelmis, Dr. Salimeh Sekeh

Keywords

averaging dynamics, learning theory, network dynamics, cascade models, PAC learning, simulations

Subject Categories

Computer Sciences

Abstract

The primary objective of this dissertation is to develop techniques for the learning-theoretic analysis of network dynamics models by formulating and studying network averaging dynamics models as a special but nontrivial case. In such models, the goal is to model the evolution of a signal on the nodes of a graph, where a node’s signal is derived from a weighted average of those of its infecting neighbors. Here the weights intuitively measure the extent to which a newly infected node “pays attention to” each of its infectors. The weighted averaging mechanism arises in domains such as network opinion dynamics, where the signal is a numerical measure of the node’s “opinion” on an issue. Such models in the literature generally treat the weights in the averages as model parameters (one per edge in the graph), and the works that study them do not focus on the problem of learning them, instead focusing on classifying asymptotic regimes of polarization, consensus, and fragmentation. Estimating all parameters thus requires a large number of identically distributed sample opinion dynamics trajectories. In this dissertation, we propose a parametric model for weighted averaging dynamics in which weights are derived from observable feature vectors on nodes via a parametric model with a small number of parameters, with the goal of addressing the limitations of traditional models.

For the proposed model class, we study several foundational statistical questions: how to derive provably consistent parameter estimators, the sample complexity of learning risk-minimizing model parameters in terms of graph and cascade structure, and identifiability of model parameters. We then show an example application of the proposed model in a downstream task: sentiment maximization. Further empirical findings from simulation studies reveal interesting insights on the role of node features in the presence of graph topology. Finally, we show how similar modeling techniques can be used for modeling of spreading processes.

Overall, this dissertation is a step toward a unified statistical toolkit for learning-theoretic aspects of network dynamics models, providing theoretical guarantees and empirical insights.

Comments

Revision (submission)

License

This work is licensed under the University at Albany Standard Author Agreement.

Download

Available for download on Saturday, May 15, 2027

Share

COinS