Maintaining Ferment: On Opinion Control Over Social Networks

Mohak Goyal, Nikhil Karamchandani, Debasish Chatterjee, D. Manjunath

Abstract

We consider the design of external inputs to achieve a control objective on the opinions, represented by scalars, in a social network. The opinion dynamics follow a variant of the discrete-time Friedkin-Johnsen model. We first consider two minimum cost optimal control problems over a finite interval $(T_0,T),$ $T_0 > 0$ -- (1) TF where opinions at all nodes should exceed a given $τ,$ and (2) GF where a scalar function of the opinion vector should exceed a given $τ.$ For both problems we first provide a Pontryagin maximum principle (PMP) based control function when the controllable nodes are specified. We then show that both these problems exhibit the turnpike property where both the control function and the state vectors stay near their equilibrium for a large fraction of the time. This property is then used to choose the optimum set of controllable nodes. We then consider a third system, MF, which is a cost-constrained optimal control problem where we maximize the minimum value of a scalar function of the opinion vector over $(T_0,T).$ We provide a numerical algorithm to derive the control function for this problem using non-smooth PMP based techniques. Extensive numerical studies illustrate the three models, control techniques and corresponding outcomes.

Introduction

In social networks, opinions often naturally decay toward a "quiescent" state—people lose interest, or they revert to their original prejudices. For someone running a campaign (whether it's for brand awareness, political mobilization, or countering misinformation), the goal is to maintain a minimum level of interest or "mindshare" across the population. This paper explores the intersection of social science and control theory, providing a mathematical framework for how an external agent can strategically influence a social network to keep a specific topic or opinion "alive" at a minimum cost.

The challenge is twofold: first, the control problem of determining how much influence to exert on specific individuals over time; and second, the design problem of selecting which specific "influencers" (nodes) to target within a large network.

Methodology

The authors model opinion dynamics using a discrete-time variant of the Friedkin-Johnsen model, where an individual's opinion is a weighted combination of their neighbors' opinions and their own internal prejudice. They address three optimization problems:

  • Total Ferment (TF): Minimize the cost to keep every person's opinion above a certain threshold.
  • Group Ferment (GF): Minimize the cost to keep a general "group opinion" above a threshold.
  • Maxmin Group Ferment (MF): Given a fixed budget, maximize the minimum level of group engagement achieved over a period.

To solve these, they employ the Pontryagin Maximum Principle (PMP). For the MF problem, which involves a non-differentiable "minimum" function, they develop a specialized non-smooth PMP numerical algorithm.

The Turnpike Property and Node Selection

The most significant theoretical finding is the turnpike property. The authors prove that for long-term campaigns, the optimal opinion states and control inputs stay very close to a specific equilibrium for the vast majority of the time. This is illustrated in the following figures, where both states and control actions quickly plateau after a brief startup period.

State trajectories showing the turnpike property. Opinions quickly reach an equilibrium and stay there until the very end of the time horizon.
Control actions showing the turnpike property. The effort exerted on controlled nodes remains constant for most of the duration.

Because of the turnpike property, the complex problem of picking which nodes to control can be simplified. Instead of looking at the whole time horizon, one can optimize for the steady-state (equilibrium) cost. The authors found that a simple greedy heuristic consistently outperformed standard network metrics like "degree centrality."

Results and Discussion

The researchers evaluated their algorithms using synthetic networks (Barabási-Albert, Erdős-Rényi, and k-regular) and the real-world Zachary’s Karate Club network. Key results include:

  • Network Structure: It is significantly cheaper to maintain ferment in Barabási-Albert (BA) networks than in regular graphs, as "hubs" efficiently spread influence.
  • Stubbornness: The cost of control increases exponentially as the "stubbornness" of the population increases.
  • Threshold Effects: The cost of maintaining ferment is approximately quadratic with respect to the desired opinion threshold ($\tau$).

Conclusion and Future Work

This work establishes a robust foundation for opinion control in social networks. While the greedy heuristic is effective, future work aims to formally prove the submodularity of the node selection problem. Additionally, the authors suggest exploring adversarial control scenarios where multiple agents compete for influence, as well as improving numerical stability for sharp opinion transitions in the MF problem.

Summary updated

Abstract

We consider the design of external inputs to achieve a control objective on the opinions, represented by scalars, in a social network. The opinion dynamics follow a variant of the discrete-time Friedkin-Johnsen model. We first consider two minimum cost optimal control problems over a finite interval $(T_0,T),$ $T_0 >0$ -- (1) TF where opinions at all nodes should exceed a given $τ,$ and (2) GF where a scalar function of the opinion vector should exceed a given $τ.$ For both problems we first provide a Pontryagin maximum principle (PMP) based control function when the controllable nodes are specified. We then show that both these problems exhibit the turnpike property where both the control function and the state vectors stay near their equilibrium for a large fraction of the time. This property is then used to choose the optimum set of controllable nodes. We then consider a third system, MF, which is a cost-constrained optimal control problem where we maximize the minimum value of a scalar function of the opinion vector over $(T_0,T).$ We provide a numerical algorithm to derive the control function for this problem using non-smooth PMP based techniques. Extensive numerical studies illustrate the three models, control techniques and corresponding outcomes.
BibTeX
@misc{goyal2019maintaining,
  author = {Mohak Goyal and Nikhil Karamchandani and Debasish Chatterjee and D. Manjunath},
  title = {{Maintaining Ferment: On Opinion Control Over Social Networks}},
  year = {2019},
  eprint = {1912.06343},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.06343},
}
Figure