General Opinion Formation Games with Social Group Membership

Vittorio Bilò, Diodato Ferraioli, Cosimo VinciIJCAI 2022 · DOI

TL;DR

This paper introduces a generalized opinion formation model that accounts for social group membership, attraction to friends, and repulsion from enemies, offering a more nuanced view of how opinions spread. The authors prove that this model always admits a stable equilibrium and provide bounds on the "price of anarchy" and "price of stability," showing that these equilibria are often close to optimal, especially when balancing truthfulness and social pressure.

What Problem Does This Paper Address?

Understanding how individuals form opinions and how social interactions influence this process is crucial for fields like marketing and political science. Existing models, such as the DeGroot and Friedkin-Johnsen models, primarily focus on agents imitating their "friends." However, real-world opinion formation is more complex. People are influenced not only by friends but also by "enemies" (leading to repulsion), and their opinions are often shaped by their membership in social groups, which can limit the influence of outsiders. Furthermore, the strength of influence from different individuals can vary significantly, even within the same social relationship. This paper addresses these limitations by proposing a more comprehensive framework that captures these multifaceted influences.

Methods

The authors introduce a generalized opinion formation game (GOF game) modeled as a cost minimization game. Each agent $i$ has a private belief $s_i$ and a public opinion $x_i$. The core of the model lies in the cost functions:

  • Private Cost: $g_i(|x_i - s_i|)$ represents the cost of deviating from one's own belief.
  • Public Influence Cost: $\sum_{j \neq i} f_{i,j}(|x_i - x_j|)$ represents the cost of social pressure. The function $f_{i,j}$ can be increasing (attraction) or decreasing (repulsion) depending on whether $j$ is a friend or enemy of $i$.

Key novelties include:

  • Social Group Membership: Agents have maximum deviation limits ($d_i^{\pm}$) from their private belief, dictated by their social group, constraining their opinions to stay within a certain range.
  • Attraction and Repulsion: The $f_{i,j}$ functions allow for both positive (friend) and negative (enemy) influence.
  • Pair-Specific Influence Strengths: The $f_{i,j}$ and $g_i$ functions can be arbitrary and pair-specific, allowing for nuanced influence dynamics.

The paper proves the existence of a pure Nash equilibrium for any GOF game. For "well-ordered" games (a specific class of convex games with non-intersecting social groups and ordered opinions), the equilibrium is unique. The authors then analyze the performance of these equilibria using the concepts of Price of Anarchy (PoA) and Price of Stability (PoS), measured by a novel "$\lambda$-social influence cost" which is a convex combination of the cost of untruthfulness and the cost of social pressure.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and analysis rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Realism through Complexity: The model's strength lies in its ability to capture more realistic social dynamics by incorporating group membership, attraction/repulsion, and varying influence strengths, moving beyond simple imitation.
  • Guaranteed Stability: Despite the added complexity, the model guarantees the existence of at least one stable opinion profile (Nash equilibrium), which is a fundamental property for any predictive model of opinion formation.
  • Balancing Truth and Social Pressure: The introduction of the $\lambda$-social influence cost provides a flexible metric to evaluate equilibria. By tuning $\lambda$, one can prioritize truthfulness ($\lambda \approx 1$), consensus ($\lambda \approx 0$), or a balance between the two ($\lambda \approx 0.5$). This offers a practical way to assess the trade-offs in opinion dynamics.
  • Bounded Suboptimality: For many realistic scenarios, especially when $\lambda$ is not at its extreme values, the equilibria are shown to be close to the optimal social outcome. This suggests that decentralized opinion formation, even with selfish agents, doesn't necessarily lead to drastically poor social performance.
  • Importance of Social Structure: The analysis hints that avoiding enemy relations and ensuring "closed" social groups (as in well-ordered games) can lead to better equilibrium outcomes, suggesting potential interventions for social planners.

Key Results

  • Equilibrium Existence: All GOF games admit at least one pure Nash equilibrium.
  • Unique Equilibrium in Well-Ordered Games: For strictly convex well-ordered games, the equilibrium is unique.
  • Price of Anarchy (PoA) Bounds:
    • The PoA can be unbounded for general unconstrained convex isolation games.
  • For well-ordered GOF games, the PoA is bounded by $\frac{\max\{2\lambda, 1-\lambda\}}{\min\{2\lambda, 1-\lambda\}}$. This bound is often tight and suggests good performance for intermediate $\lambda$ values.
  • Price of Stability (PoS) Bounds:
    • For any GOF game, PoS is bounded by $\frac{\max\{2\lambda, 1-\lambda\}}{\min\{2\lambda, 1-\lambda\}}$.
  • Matching lower bounds are provided for specific cases, demonstrating the tightness of these bounds.
  • Practical Implications: The bounds suggest that for $\lambda$ values not close to 0 or 1, the performance of equilibria is good, indicating that agents' decentralized decisions can lead to socially acceptable outcomes.

Limitations and Future Work

  • Computational Complexity: While equilibrium existence is proven, finding these equilibria or analyzing their properties can be computationally challenging for large networks.
  • Assumptions for Bounds: Some of the tighter bounds on PoA and PoS rely on specific assumptions like convexity and the "well-ordered" game structure, which might not hold in all real-world scenarios.
  • Future Work: The authors suggest embedding the opinion formation process in an evolving setting to study radical changes in cultures and styles. They also mention the potential for applying the model to analyze marketing strategies (e.g., targeted advertising) and election control. Further exploration of the primal-dual method for deriving bounds in more general settings is also a promising direction.

Summary updated

BibTeX
@inproceedings{bilo2022general,
  author = {Vittorio Bilò and Diodato Ferraioli and Cosimo Vinci},
  title = {{General Opinion Formation Games with Social Group Membership}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/13},
}
Figure