Opinion Dynamics with Local Interactions

Dimitris Fotakis, Dimitris Palyvos-Giannas, Stratis Skoulakis

Abstract

Traditional models of opinion dynamics, such as the Friedkin-Johnsen (FJ) and Hegselmann-Krause (HK) models, typically assume that agents have access to global information about the population. This paper investigates the behavior of these models under more realistic constraints: limited information exchange and interactions restricted by fixed social networks. The authors provide theoretical proofs for the convergence of these local variants and use extensive simulations to characterize their behavior. Key findings indicate that while random sampling can effectively mimic global dynamics, fixed social structures often lead to significant opinion fragmentation and the formation of numerous isolated "singleton" clusters.

Introduction

Understanding how opinions evolve and stabilize within a population is a central challenge in social sciences and multi-agent systems. Mathematical frameworks like the Friedkin-Johnsen (FJ) and Hegselmann-Krause (HK) models have long been used to study these dynamics. However, a major limitation of these classical models is the assumption of "perfect information"—the idea that every agent can interact with every other agent in every time step.

In the real world, especially on large-scale social media platforms, such global interaction is impossible. Individuals typically interact with a small subset of their social circle or encounter random snippets of information. This paper addresses this gap by exploring how opinion dynamics behave when interactions are restricted to local neighborhoods or random samples.

Modeling Local Interactions

The researchers introduce three primary variants to study local dynamics:

  1. Limited Information Protocol for FJ (LIP-FJ): In the standard FJ model, agents update their opinions by averaging the views of all their neighbors. In LIP-FJ, an agent selects exactly one neighbor at random in each round. The authors prove that this randomized process converges in expectation to the same unique stable state as the standard model, though the convergence rate is significantly slower.
  2. Network-HK: This variant combines bounded confidence with a fixed social graph. Agents only consider the opinions of neighbors who are both connected to them in a social network and whose opinions fall within a specific "confidence bound" (epsilon).
  3. Random-HK: Instead of a fixed network, agents sample a small random subset of the entire population in each round. They then update their opinions based only on those in the sample who fall within their confidence bound.

Convergence and Fragmentation

A significant contribution of this work is the theoretical proof of convergence for these local HK variants. By utilizing the concept of temporal network connectivity and the coefficient of ergodicity for stochastic matrices, the authors demonstrate that these systems will eventually reach a stable equilibrium, even with asymmetric interactions.

Simulations reveal a fascinating "peak" in convergence time. As the confidence bound epsilon increases, the time required for the population to reach consensus or stable clusters initially increases, peaking at a phase transition point before decreasing.

Comparison of convergence times for standard HK and Random-HK (sampling) models, highlighting the peak convergence time near the phase transition point.

The Impact of Social Networks

The most striking differences appear when comparing global models to those restricted by a fixed social network. In the standard HK model, populations typically coalesce into a few large opinion clusters. However, in the Network-HK model, the social graph acts as a filter that prevents similar-minded individuals from interacting if they are not directly connected.

This restriction leads to cluster proliferation. For small values of epsilon, a population that would normally form two or three large groups instead fragments into hundreds of "singleton" clusters—individuals who become isolated because no one in their immediate social circle shares a sufficiently similar view.

Comparison of final cluster counts between standard HK and Network-HK (graph) models. The fixed network structure leads to extreme fragmentation at low confidence bounds.

On real-world networks, such as the Facebook "circles" dataset from SNAP, these effects are even more pronounced. Tightly bound communities act as information silos, where internal communication is frequent but "bridge" edges to other groups are scarce, creating significant bottlenecks for information flow and slowing down the overall convergence of the system.

Conclusion

This research demonstrates that while local and randomized interactions do not prevent opinion models from converging, they fundamentally alter the resulting social landscape. Random sampling can serve as a proxy for global interaction if the sample size is sufficiently large. However, fixed social networks impose structural constraints that lead to increased polarization and fragmentation. These insights are crucial for understanding the dynamics of modern digital communication environments, where algorithmic filtering and social graph constraints play a dominant role in shaping public opinion.

Summary updated

Abstract

We study convergence properties of opinion dynamics with local interactions and limited information exchange. We adopt a general model where the agents update their opinions in rounds to a weighted average of the opinions in their neighborhoods. For fixed neighborhoods, we present a simple randomized protocol that converges in expectation to the stable state of the Friedkin-Johnsen model. For opinion-dependent neighborhoods, we show that the Hegselmann-Krause model converges to a stable state if each agent's neighborhood is restricted either to a subset of her acquaintances or to a small random subset of agents. Our experimental findings indicate that for a wide range of parameters, the convergence time and the number of opinion clusters of the neighborhood-restricted variants are comparable to those of the standard Hegselmann-Krause model.
BibTeX
@inproceedings{fotakis2016opinion,
  author = {Dimitris Fotakis and Dimitris Palyvos-Giannas and Stratis Skoulakis},
  title = {{Opinion Dynamics with Local Interactions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure