Learning to Control Misinformation: a Closed-loop Approach for Misinformation Mitigation over Social Networks

TL;DR

The paper tackles misinformation amplification in social networks caused by engagement-optimized recommender systems. It extends the closed-loop Friedkin-Johnsen framework to model sentiment propagation and introduces a penalized cost $\theta_{\rm M}$ that discourages extreme negativity and novelty in content, balancing misinformation mitigation with user engagement. The authors present both model-free and model-based control strategies, prove convergence, and validate the approach on LIAR2-derived sentiment features, reporting up to $76\%$ reduction in misinformation and revealing favorable engagement trade-offs in radicalized networks. This work provides actionable insights for platform operators seeking to suppress misinformation without sacrificing engagement and sets the stage for further real-world validation and adaptive control strategies.

Abstract

Modern social networks rely on recommender systems that inadvertently amplify misinformation by prioritizing engagement over content veracity. We present a control framework that mitigates misinformation spread while maintaining user engagement by penalizing content characteristics commonly exploited by false information, specifically, extreme negative sentiment and novelty. We extend the closed-loop Friedkin-Johnsen model to incorporate the mitigation of misinformation together with the maximization of user engagement. Both model-free and model-based control strategies demonstrate up to 76% reduction in misinformation propagation across diverse network configurations, validated through simulations using the LIAR2 dataset with sentiment features extracted via large language models. Analysis of engagement-misinformation trade-offs reveals that in networks with radical users, median engagement improves even as misinformation decreases, suggesting content moderation enhances discourse quality for non-extremist users. The framework provides practical guidance for platform operators in balancing misinformation suppression with engagement objectives.

Figures (3)

• Figure 1: Sentiment Dynamics Evolution. Top: 100-User Network showing mean user emotional extremity (solid) and recommender output (dashed) over 100 time steps, comparing baseline engagement-only control ($\theta$, green) with MF mitigation ($\theta_{\rm M}$, blue) and MB mitigation ($\theta_{\rm M}$, orange) at $\rho=2.5$. Left: synthetic continuous dynamics; Right: data-driven discrete content. Shaded regions indicate standard deviation. Bottom: 6-User Network with Radical User showing the same comparison over 50 time steps. Individual user trajectories are shown in light lines. The stubborn radical user remains at maximum emotional extremity (top of plot), while mitigation strategies prevent negativity propagation to other users.
• Figure 2: Performance Metrics Analysis for 100-User Network (Data-Driven). Top-left: Misinformation metric $\mathcal{M}$ vs. penalty weight $\rho$. Top-right: Per-user engagement cost vs. $\rho$ (mean and median). Bottom-left: Sentiment shift $|x_i(\tau) - x_i(0)|$ vs. $\rho$ (mean and median). Bottom-right: Pareto frontier showing trade-off between median per-user engagement cost and misinformation (lower-left is better). Labeled points indicate $\rho$ values. Blue: Model-Free; Orange: Model-Based.
• Figure 3: Performance Metrics Analysis for 6-User Radical Network (Data-Driven). Top-left: Misinformation metric $\mathcal{M}$ vs. $\rho$. Top-right: Per-user engagement cost vs. $\rho$; note that median (dashed) remains stable while mean (solid) increases, indicating improved engagement for non-radical users. Bottom-left: Sentiment shift vs. $\rho$. Bottom-right: Pareto frontier using median engagement, showing that mitigation can improve both objectives simultaneously in radicalized networks. Blue: Model-Free; Orange: Model-Based.