Social Learning under Platform Influence: Consensus and Persistent Disagreement

Ozan Candogan, Nicole Immorlica, Bar Light, Jerry AnunrojwongarXiv 2022 · DOI · arXiv

TL;DR

This paper explores how social media platforms—driven by engagement-maximizing algorithms—interfere with the natural human tendency to reach consensus through social learning. By modeling opinion dynamics on large networks, the authors prove that complex social systems can be simplified into a "two-agent" approximation, revealing that platform influence is the primary driver of persistent disagreement and political polarization.

What Problem Does This Paper Address?

In classical "DeGroot" models of social learning, individuals in a connected network eventually reach a consensus by averaging the opinions of their neighbors. However, our modern reality is characterized by deep, persistent polarization. Why doesn't the internet bring us together?

The authors argue that the missing piece is the platform. Social networking sites don't just facilitate peer-to-peer talk; they use algorithms to maximize engagement. Because users engage more with content that matches their existing views, platforms effectively act as a "neighbor" that constantly nudges users toward extreme versions of their own beliefs. The paper seeks to mathematically characterize the "tug-of-war" between social learning (which pulls people together) and platform influence (which pushes them apart).

Methods

The researchers introduce a naive social learning model where an agent’s opinion evolves based on two factors:

  1. Social Influence: The weighted average of their neighbors' opinions.
  2. Platform Influence: Personalized content recommendations that match the agent's current stance (modeled using a continuous interpolation of a sign function).

They primarily analyze this through a Stochastic Block Model (SBM) with two blocks (e.g., "Left" and "Right"). Their most significant theoretical contribution is a mean-field approximation theorem. They prove that for large, dense networks, the collective behavior of thousands of agents can be accurately modeled by a simple two-agent system. This reduction allows them to derive tractable analytical conditions for when a society will reach consensus versus when it will remain polarized. They also use Lyapunov functions to prove that under symmetric influence, opinions will always converge to a stable equilibrium rather than oscillating forever.

Datasets

While the paper is primarily theoretical, the authors validate their model using:

  • Synthetic Data: Simulations of Stochastic Block Models with varying numbers of agents ($n=60$ to $300$) and different connection probabilities.
  • Real-World Data: The Political Blogs dataset (Adamic and Glance, 2005), which contains 1,222 nodes and over 33,000 edges representing links between liberal and conservative blogs during the 2004 US election.

Key Insights and Ideas

  • The Tug-of-War: Consensus is a result of social learning winning the battle. Persistent disagreement happens when the platform's "nudge" is strong enough to overcome the pull of one's neighbors.
  • The "Moderate" Sweet Spot: The authors find that "extremism" (how far opinions are from neutral) is non-monotonic. When a platform is very weak, you get consensus at an extreme (everyone agrees on one side). When it's very strong, you get extreme disagreement. A "moderate" platform actually leads to the lowest overall extremism by allowing for moderate disagreement.
  • Initial Conditions Matter: Whether a group reaches consensus isn't just about the algorithm; it's about the starting point. "Balanced" initial opinions (where the average is near neutral) are actually more likely to lead to persistent disagreement because neither side has enough "gravitational pull" to convert the other.
  • Echo Chambers by Design: When the authors extended the model to include a correlation between initial opinions and connection probabilities (homophily), they found it amplified final polarization, as agents predominantly interact with peers who reinforce the platform's bias.

Key Results

  • Approximation Accuracy: Simulations showed that the mean equilibrium opinions in large networks quantitatively match the predictions of the simplified two-agent system.
  • Polarization Monotonicity: The paper proves that polarization is monotonic along the trajectory; if you start more polarized than the eventual equilibrium, polarization will steadily decrease until it hits that limit (and vice versa).
  • Consensus Thresholds: The authors provide specific mathematical inequalities (Conditions C1 and C2) that define the boundaries between consensus and disagreement based on platform strength ($b$) and social tie strength ($a$).
  • Three-Block Dynamics: In a model with a "centrist" media outlet, increasing the "moderation parameter" (the range of opinions the platform considers neutral) directly reduces final system-wide polarization.

Limitations and Future Work

  • Symmetric Influence: The proof for convergence to an equilibrium relies on a symmetric influence matrix (agent A influences B as much as B influences A). The authors show via a 4-agent counterexample that without this, opinions can enter limit cycles (perpetual oscillation).
  • Descriptive vs. Strategic: The model assumes agents are "naive" and follow heuristic rules. It does not account for strategic behavior, where agents might intentionally change their stance to gain followers or avoid certain content.
  • Lack of "Truth": The model deals purely with opinions, not facts. It doesn't account for "ground truth" or the spread of misinformation/fake news, which are critical components of modern platform influence.
  • Welfare Analysis: The authors note that while their model describes what happens, it doesn't yet tackle the "social good" aspect—designing algorithms that specifically aim to minimize polarization while maintaining engagement.

Summary updated

Abstract

Individuals increasingly rely on social networking platforms to form opinions. However, these platforms typically aim to maximize engagement, which may not align with social good. In this paper, we introduce an opinion dynamics model where agents are connected in a social network, and update their opinions based on their neighbors' opinions and on the content shown to them by the platform. We focus on a stochastic block model with two blocks, where the initial opinions of the individuals in different blocks are different. We prove that for large and dense enough networks the trajectory of opinion dynamics in such networks can be approximated well by a simple two-agent system. The latter admits tractable analytical analysis, which we leverage to provide interesting insights into the platform's impact on the social learning outcome in our original two-block model. Specifically, by using our approximation result, we show that agents' opinions approximately converge to some limiting opinion, which is either: consensus, where all agents agree, or persistent disagreement, where agents' opinions differ. We find that when the platform is weak and there is a high number of connections between agents with different initial opinions, a consensus equilibrium is likely. In this case, even if a persistent disagreement equilibrium arises, the polarization in this equilibrium, i.e., the degree of disagreement, is low. When the platform is strong, a persistent disagreement equilibrium is likely and the equilibrium polarization is high. A moderate platform typically leads to a persistent disagreement equilibrium with moderate polarization. We analyze the effect of initial polarization on consensus and explore various extensions including a three block stochastic model and a correlation between initial opinions and agents' connection probabilities.
BibTeX
@misc{candogan2022social,
  author = {Ozan Candogan and Nicole Immorlica and Bar Light and Jerry Anunrojwong},
  title = {{Social Learning under Platform Influence: Consensus and Persistent Disagreement}},
  year = {2022},
  eprint = {2202.12453},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.12453},
}
Figure