Abstract

This dissertation presents five main contributions to the field of opinion dynamics.

The first contribution introduces a discrete opinion-dynamics model for social-influence networks that incorporates both synchronous and asynchronous updating mechanisms. Each agent holds two distinct states — a synchronously updated private opinion and an asynchronously updated expressed opinion. Private opinions evolve through interactions with neighbors’ expressed opinions, while expressed opinions are shaped by neighboring views and pressure to conform to public discourse. The model is analyzed under several network topologies, and simulation results confirm its ability to capture realistic patterns of online opinion formation.

The second contribution is the collection, preprocessing, and public release of a large-scale Twitter/X dataset spanning four months of discussion on a focused topic. The dataset includes opinions, timestamps, geo-tags, stance labels, and follower–friend graphs. Thorough cleaning, bot removal, and language filtering ensure a research-ready corpus that supports longitudinal studies of network structure and opinion evolution.

The third contribution analyzes competitive influence in social-media networks by combining the evolved DeGroot-based model with reinforcement learning. Agents employ Q-learning to decide when to voice their views, aiming to maximize their impact on connected individuals. By varying topologies and examining convergence behaviors, the study reveals how strategic timing affects polarization and consensus formation, underscoring the role of competition in shaping discourse.

The fourth contribution extends the reinforcement-learning framework by formalizing the opinion-dynamics environment as a partially observable Markov decision process and introducing adaptive control agents whose policies update at every time step. Robustness tests across networks of different sizes, densities, and topologies demonstrate that learned policies can steer collective opinion toward arbitrary targets with minimal intervention. Open-source code and cross-network experiments validate the scalability and generalizability of this approach.

The fifth contribution develops a Double Deep Q-Learning controller that strategically amplifies polarization and disagreement in online networks. Integrating the controller with the Expressed-and-Private Opinion model, the study shows that manipulating only a small subset of accounts can markedly increase ideological division once a topology-specific threshold is crossed. These findings shed light on how algorithmic interventions can engineer polarization in digital spaces.

The thesis concludes with a discussion of the limitations and directions for future research.