The contemporary information environment is increasingly shaped by large-scale online influence campaigns that exploit digital platforms to manipulate public opinion, amplify polarization, and erode trust in institutions. While synthetic population models and agent-based simulations have long been used to study social behavior, existing approaches struggle to fully represent the cognitive, emotional, and narrative mechanisms that drive belief formation and opinion change in modern digital ecosystems. Classical opinion dynamics models often assume psychologically homogeneous agents, static influence structures, and topic-agnostic interactions. These assumptions limit their applicability to social-cyber systems, where human discourse is deeply interwoven with algorithmically generated content.

This dissertation introduces an integrated, AI-driven framework, for synthetic populations and opinion dynamics modeling designed to simulate social-cyber influence in the online information environment. At its foundation, is a synthetic population modeling pipeline that constructs demographically grounded skeleton personas using empirical population data and statistical synthesis methods. These skeleton personas preserve population-level realism and internal consistency, while remaining intentionally sparse. This emables subsequent enrichment without sacrificing transparency, reproducibility, or control.

This framework introduces the use of AURORA (AI-Utilized Retrieval for Optimized Representation of Audiences), a Retrieval-Augmented Generation (RAG–based persona modeling system, to enrich skeleton personas with psychologically and contextually realistic attributes. AURORA leverages large language models (LLMs), semantic vector databases, and salience-aware retrieval to assign emotional traits, ideological orientations, topic-specific salience, and narrative grounding to individual agents. This design enables the generation of internally coherent synthetic populations whose beliefs, emotions, and attention patterns are grounded in real-world discourse while remaining modular and configurable for experimental and operational use.

Building on these enriched populations, this dissertation presents JANUS (Joint Agent Network for Understanding Social Dynamics). JANUS is an extension of classical opinion dynamics models that incorporates emotion, salience, and persuasion into a mechanistic and interpretable update framework. JANUS extends the Friedkin–Johnsen model by introducing heterogeneous susceptibility, identity-conditioned signed influence, engagement probability, and affective drift. These extensions enable topic-dependent opinion evolution within psychologically diverse populations. Empirical evaluations demonstrate that the proposed framework sustains heterogeneous and polarized opinion states, without collapsing to consensus. It also reproduces empirically observed emotion and salience driven behaviors. Together, AURORA and JANUS form a unified foundation for scalable simulation of social-cyber influence, advancing the state of the art in synthetic population modeling, opinion dynamics, and information environment analysis.

343 pages

Thesis Committee:
Kathleen M. Carley (Chair)
Fei Fang
Rick Carley
Johnathon Harris (Office of Naval Research)

Nicolas Christin, Head, Software and Societal Systems Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu