Harmonizing vs Polarizing Platform Influence Functions

Hind AlMahmoud, Frederik Mallmann-TrennarXiv 2024 · DOI · arXiv

Abstract

This paper introduces a mathematical framework to analyze the impact of algorithmic content curation on social opinion dynamics. By modeling the interplay between social influence and platform-driven "forces" using ordinary differential equations, the authors identify the conditions under which algorithms lead to societal consensus or persistent polarization. Central to this work is the introduction of "polarization planes," a geometric tool that characterizes steady states in opinion evolution. The findings demonstrate that specific platform functions, such as those reinforcing extreme views, can mathematically guarantee a divided society, while moderating functions can foster global agreement.

Introduction

The role of social media algorithms in shaping public opinion has become a central concern in modern discourse. While traditional models of social learning focus on how individuals influence one another, they often overlook the "tug-of-war" between peer influence and algorithmic personalization. This research addresses this gap by asking: Can we mathematically predict whether a platform's design is "harmonizing" (promoting agreement) or "polarizing" (entrenching division)?

Methodology: Modeling Opinion Dynamics

The authors model opinion dynamics through an Ordinary Differential Equation (ODE) where an individual's opinion changes over time based on two primary factors:

  1. Peer Influence: A weighted average of opinions from a user's social network, following the classic DeGroot model.
  2. Platform Influence: A force pulling the user's opinion toward a target curated by the platform’s influence function, $f(x)$.

To analyze these dynamics, the researchers introduce "polarization planes." This geometric method allows for the visualization and mathematical identification of steady states where opinions may become "stuck."

Figure 1: Geometric representation of polarization planes used to identify steady-state opinion equilibria.

The study extends these models from simple two-agent systems to complex structures, including complete graphs and Stochastic Block Models (SBMs). Through mean-field approximations, the authors demonstrate that the behavior of large-scale networks can be effectively predicted by the dynamics of smaller, representative systems.

Key Insights: The Geometry of Polarization

The core finding is that polarization occurs when the platform's influence function $f(x)$ intersects the polarization planes in specific quadrants. If an algorithm pushes a user toward an extreme faster than their social circle pulls them toward the center, consensus becomes impossible.

Figure 3: Analysis of the intersection between platform influence functions and polarization planes, illustrating regions of persistent disagreement.

The "Sign Function" Trap

The authors analyze a continuous version of the "sign function"—an algorithm that consistently presents content slightly more extreme than the user's current view. This design leads to five steady states: three representing consensus (at -1, 0, or +1) and two representing "persistent disagreement," where the population remains perfectly split.

Moderation and the "Anti-Sign" Function

Conversely, a "Linear Harmonizing Function" ($f(x) = \alpha x$ where $\alpha < 1$) acts as a moderating force, eventually overcoming social friction to reach a global consensus. Interestingly, an "Anti-Sign" algorithm—showing users content they disagree with—effectively destroys polarization, though it may risk user disengagement.

Results and Theoretical Contributions

  • Consensus Conditions: Strong consensus (where all agents hold identical opinions) requires the platform function to satisfy $f(y) = y$.
  • Persistent Disagreement: The paper provides necessary and sufficient conditions for polarization. If the platform function remains "above" the polarization plane for positive opinions and "below" it for negative ones, division is guaranteed.
  • Scalability to SBMs: In dense Stochastic Block Models, the opinions of "left" and "right" agent blocks converge to values predictable by a two-person model, validating the utility of small-scale modeling for large social behaviors.

Limitations and Future Directions

While the framework identifies equilibrium states, it does not fully map "regions of attraction"—the specific starting conditions that determine which equilibrium a system will reach. Additionally, the current model assumes symmetric influence between users, whereas real-world networks often feature asymmetric power dynamics (e.g., influencers). Future research may also explore the economic trade-off between platform engagement (often driven by polarization) and societal health.

Summary updated

BibTeX
@misc{almahmoud2024harmonizing,
  author = {Hind AlMahmoud and Frederik Mallmann-Trenn},
  title = {{Harmonizing vs Polarizing Platform Influence Functions}},
  year = {2024},
  eprint = {2410.09225},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.09225},
}
Figure