Back to search

Boston University

Statistical learning for dynamic networks

Abstract

dc:description.abstract

We consider the problem of modeling dynamic networks, a collection of network graphs indexed over times. Latent space network models are a commonly used class of models where the probability of a relation between actors depends on the positions of individuals in an unobserved latent space. In the Coevolving Latent Space Network with Attractors (CLSNA) model, temporal evolution of latent positions depends on network connectivity via the presence of attractors. This dissertation advances the CLSNA model along three directions. The first project addresses the scalability and flexibility limitations of the original CLSNA model. We develop an SGD parameter estimation method that allows us to scale to larger networks, accompany that by a novel corresponding approach to variance estimation that builds on the notion of Laplace approximation, and implement these advances within an extension of the CLSNA framework that allows nodes to join and leave the network. Application to the Congressional Hashtag Network on X demonstrates that accounting for node dynamics overcomes selection bias and uncovers previously concealed and increasingly repulsive forces within the Republican Party. The second project significantly extends the CLSNA model by introducing dyadic covariates into the link probability function and applying the framework for the first time to a multi-party system. We introduce party affiliation, gender, and age gap as fixed dyadic covariates, so that the dynamically updating latent positions capture residual ideological proximity rather than demographic structure. In addition, we replace the binary attraction matrix with a full symmetric matrix, allowing pairwise forces across Left, Center, and Right blocs to be estimated separately. The third project explores the consistency of estimators within dynamic network models with stochastic attractors. We focus on a subset of the general CLSNA class and consider three cases of increasing complexity: attraction to the origin, attraction to the global average, and attraction to random neighbors. We propose a two-step procedure combining low-rank matrix completion with linear regression, and establish that both the latent positions and the attraction parameter are consistently estimated in all three cases, providing the first formal consistency results for this class of models.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mathematics & Statistics
Grantor
Boston University
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Pan, Hancong
Advisors dc:contributor.advisor
  • Kolaczyk, Eric D.
  • Spiliopoulos, Konstantinos

Subjects

dc:subject × 1

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53615

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Pan, Hancong. Statistical learning for dynamic networks. doctoral thesis, Boston University, 2026. https://hdl.handle.net/2144/53615