Abstract
This paper studies the evolution of self-appraisal and social power for a group of individuals who discuss and form opinions. We consider a modification of the recently proposed DeGroot-Friedkin (DF) model, in which the opinion formation process takes place on the same timescale as the reflected appraisal process; we call this new model the single-timescale DF model. We provide a comprehensive analysis of the equilibria and convergence properties of the model for the settings of irreducible and reducible influence networks. For irreducible influence networks, the single-timescale DF model predicts that social power ranking is asymptotically equal to centrality ranking, with power accumulating at the top of the hierarchy. For reducible influence networks, the model behaves differently from the original DF model: individuals can retain social power if the initial conditions allow, and multiple sinks can lead to various partitions of social power at equilibrium.
Introduction
In social psychology and network science, understanding how opinions change and how an individual's influence (social power) evolves is crucial. The popular DeGroot-Friedkin (DF) model typically assumes a "separation of timescales": a group discusses an issue until consensus is reached, and only then do individuals update their self-appraisal for the next issue. This paper challenges that assumption by proposing a single-timescale model where social power and opinions evolve simultaneously. This approach is arguably more realistic for real-world discussions where status and influence shift mid-debate.
Methods and Network Topologies
The authors define a dynamical system where the self-weight (social power) of an individual at the next step is a function of the current influence network. They use a row-stochastic matrix $W$ to represent influence, with the update rule $x(t + 1) = W(x(t))^T x(t)$, where $x$ represents the vector of social power.
The paper analyzes two main types of networks:
- Irreducible: Strongly connected networks where everyone can eventually influence everyone else.
- Reducible: Networks with "sinks" or subgroups that listen to each other but ignore outsiders.
Datasets and Validation
The mathematical findings were validated using numerical simulations on classic social network datasets, including Sampson’s Monastery Network and Krackhardt’s Advice Network. The latter represents the advice-seeking relationships among 21 managers in a manufacturing organization.
Key Insights and Results
Timescale Robustness in Connected Networks
In strongly connected (irreducible) networks, the timescale shift does not change the final outcome. Both models predict that social power will eventually align with eigenvector centrality. Essentially, your power is determined by the power of the people who listen to you. In star networks, this leads to the inevitable emergence of an autocrat who captures 100% of the social power.
Sensitivity in Reducible Networks
The most significant departure from the original model occurs in reducible networks with multiple "sinks." While the original DF model converges to a single global equilibrium, the single-timescale model allows for any partition of power among the sinks, meaning the final distribution is highly dependent on initial conditions. Furthermore, individuals who do not influence others (reducible nodes) can actually retain their power in this model if they start with it, unlike in the original model where their power asymptotically vanishes.
Simulations show that the single-timescale model is generally less "smooth" and takes more iterations to reach equilibrium compared to the original two-timescale approach, as seen in the extended iteration counts in the plots above.
Conclusion and Future Work
The single-timescale DF model provides a more nuanced view of social power evolution, particularly in fragmented networks. Future research could expand this by incorporating "stubborn individuals" who anchor their opinions to initial beliefs or by exploring alternative dynamics like the Friedkin-Johnsen model to see if these single-timescale assumptions hold across different psychological frameworks.
Summary updated