Pacing Opinion Polarization via Graph Reinforcement Learning

This paper introduces PACIFIER, a graph reinforcement learning framework for sequential polarization moderation on social networks. It formulates the canonical ModerateInternal (MI) and ModerateExpressed (ME) problems as sequential decision-making tasks, enabling scalable and adaptive intervention policies without relying on repeated steady-state recomputation. The framework is objective-agnostic and naturally extends to cost-aware, nonlinear, and topology-altering settings. A temporal-aware node marking mechanism and polarization-aware global features are proposed to address representation challenges. Extensive experiments on 15 real-world Twitter networks (up to 155,599 nodes) demonstrate that PACIFIER consistently outperforms non-learning baselines across multiple tasks, establishing graph RL as a unified and robust paradigm for moderating opinion polarization. Version: v1 (2025-04-01). This is a preprint.