Analysis of hierarchically polarized communities on social media

Sun, Dachun

Permalink

https://hdl.handle.net/2142/130019 Copy

Description

Title

Analysis of hierarchically polarized communities on social media

Author(s)

Sun, Dachun

Issue Date

2025-07-08

Director of Research (if dissertation) or Advisor (if thesis)

Abdelzaher, Tarek

Doctoral Committee Chair(s)

Abdelzaher, Tarek

Committee Member(s)

• Tong, Hanghang
• Ji, Heng
• Lebiere, Christian

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Social Network Analysis
• Polarization Identification
• Community Response Generation

Abstract

Social media platforms have profoundly transformed public discourse, providing a global forum for individuals to share information, express opinions, and build communities. While they foster unprecedented connectivity, they also exacerbate polarization, forming communities characterized by hierarchical ideological divisions. Many existing computational approaches oversimplify these dynamics, reducing polarization to binary oppositions and neglecting the nuanced, evolving structure of beliefs within online networks. Furthermore, there has been limited exploration into predicting community responses to new hypothetical posts, leaving critical gaps in the understanding of and strategies for mitigating polarization. This dissertation addresses three central research questions: How can hierarchically polarized communities be identified and characterized on social media platforms? How can the collective responses of these communities to new social stimuli be realistically simulated, capturing their structural complexity and the cognitive mechanisms that drive opinion formation? And how can the discovered solutions be implemented in a practical tool? To answer these questions, a suite of computational techniques is developed to detect, represent, and simulate complex community structures in social media environments. A dynamic polarized belief representation framework grounded in recurrent graph autoencoders is introduced, allowing for the tracking and analysis of hierarchical community structures as they evolve over time. To address the scarcity of labeled data, a perturbation-based active learning strategy is proposed to optimize label efficiency in semi-supervised settings by strategically selecting and labeling informative nodes within the social graph. This dissertation also advances the state of the art in community response generation by developing a retrieval-augmented generation (RAG) framework that leverages both historical social media responses and external knowledge to forecast community reactions to hypothetical posts. To better reflect human biases, cognitive mechanisms such as memory recency, frequency, and similarity weighting are incorporated, emulating human biases in opinion dynamics and bridging the gap between rational computation and real-world behavior. Much of this work is integrated into an analytics tool for conflict monitoring and intervention, validated through demonstrations on real-world social media datasets. Overall, this dissertation explores these core issues related to identifying, understanding, and simulating the evolution and responses of hierarchical and dynamic polarized communities on social media, thereby supporting efforts to address the challenges of polarization and radicalization in online environments.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/130019

Copyright and License Information

Copyright 2025 Dachun Sun

Owning Collections