Memory-induced mechanism for self-sustaining activity in networks.

Armen E. Allahverdyan, Greg Ver Steeg, Aram GalstyanarXiv 2017 · DOI · OpenReview

TL;DR

This paper proposes a novel mechanism for sustained collective activity in networks, driven by agent memory rather than network loops, which are rare in social media. They demonstrate that temporal memory allows nodes to form synchronized clusters, leading to self-sustaining patterns even in tree-like structures, and that this behavior is robust to noise and can be enhanced by weighted connections.

What Problem Does This Paper Address?

The paper tackles the challenge of understanding how collective activity can persist in networks, particularly in systems like social media where traditional models relying on long feedback loops (cycles) are less applicable. Existing models often explain sustained activity through the presence of "reverberators" – cycles in the network that allow signals to propagate indefinitely. However, empirical studies of social media networks suggest they often lack these long loops, resembling tree-like structures. This raises the question: how can collective activity be sustained in such networks? The authors propose that temporal memory within individual agents, rather than network topology alone, can provide the necessary mechanism for self-sustaining patterns.

Methods

The authors introduce a simplified "integrate-and-fire" (IF) model adapted for social agents. Each agent has an "information potential" that accumulates over time. When this potential exceeds a threshold, the agent "fires" (e.g., posts content). Crucially, the model incorporates:

  • Temporal Memory ($\kappa < 1$): Agents retain a fraction of their potential from the previous time step, preventing immediate reset after firing. This is a key departure from memoryless models.
  • Network Structure: The primary network topology studied is the Cayley tree, a tree-like structure with a core-periphery organization. Triangles (loops of length 3) are also added to explore the impact of short loops.
  • Behavioral Noise: Two models of noise are introduced to account for random fluctuations in agent behavior, making the model more realistic.
  • Weighted Connections: The model is extended to include randomly weighted connections, representing varying levels of "attention" agents pay to their neighbors.

The core of their analysis involves simulating this model on the Cayley tree and observing the emergent dynamics, particularly the conditions under which collective activity is sustained.

Datasets

Not discussed in the paper. The study relies on simulated data generated by their model.

Key Insights and Ideas

  • Memory as a Sustaining Mechanism: The most significant insight is that temporal memory in individual agents can drive self-sustaining collective activity, even in the absence of long network loops. This offers an alternative explanation for persistent activity in sparse networks.
  • Clustering and Synchronization: Memory-induced activity leads to the formation of synchronized clusters of agents. These clusters exhibit partial synchronization, with agents within a cluster firing at similar frequencies.
  • Core-Periphery Dynamics: The Cayley tree's inherent core-periphery structure influences the clustering dynamics, with distinct clusters often emerging in these regions.
  • Thresholds for Sustained Activity: The model exhibits distinct upper ($Q^+$) and lower ($Q^-$) thresholds for sustained activity. Between these thresholds, the system's behavior is highly sensitive to initial conditions.
  • Robustness to Noise: Weak behavioral noise does not destroy the clustering and synchronization, but rather leads to a more robust and less initial-condition-dependent fragmentation into clusters.
  • Attention Enhances Activity: Randomly weighted connections (representing attention) facilitate sustained activity and can even lower the required coupling strength.

Key Results

  • Memory-driven clusters: For $K \ge 2$ and $\kappa < 1$ (with $u > r/\kappa$), agents form synchronized clusters, leading to self-sustaining collective activity.
  • Thresholds $Q^+$ and $Q^-$: Numerical simulations reveal distinct thresholds for sustained activity. For $q > Q^+$, activity is sustained for almost all initial states. For $q < Q^-$, activity decays to zero. In the range $Q^- < q < Q^+$, the outcome depends on initial conditions.
  • Noise-induced robustness: With weak noise, the fragmentation into clusters becomes less dependent on initial conditions, and the activity sustaining is more predictable.
  • Weighted networks: Introducing random weights (attention) facilitates sustained activity and can lower the thresholds $Q^+$ and $Q^-$. For example, they found $Q_F < Q^+$ and $Q_F < Q^-$.

Limitations and Future Work

  • Model Simplification: The study uses a simplified model of agent behavior and network structure (Cayley tree with added triangles). Real social networks are more complex.
  • Lack of Direct Data Fitting: The authors acknowledge that the model is not directly fitted to empirical data, but rather suggests directions for future empirical research.
  • Facilitating Connections Only: The model assumes all connections are "facilitating" (positive influence). The role of "inhibitory" connections, known to be important in biological neurons, is not explored.
  • Future Work: The authors suggest investigating the presence and impact of inhibitory connections in social media interactions, exploring more realistic network structures, and conducting direct data-fitting to validate their findings. They also highlight the potential for this model to explain phenomena like attention and memory in social systems.

Summary updated

Abstract

We study a mechanism of activity sustaining on networks inspired by a well-known model of neuronal dynamics. Our primary focus is the emergence of self-sustaining collective activity patterns, where no single node can stay active by itself, but the activity provided initially is sustained within the collective of interacting agents. In contrast to existing models of self-sustaining activity that are caused by (long) loops present in the network, here we focus on tree--like structures and examine activation mechanisms that are due to temporal memory of the nodes. This approach is motivated by applications in social media, where long network loops are rare or absent. Our results suggest that under a weak behavioral noise, the nodes robustly split into several clusters, with partial synchronization of nodes within each cluster. We also study the randomly-weighted version of the models where the nodes are allowed to change their connection strength (this can model attention redistribution), and show that it does facilitate the self-sustained activity.
BibTeX
@misc{allahverdyan2017memory,
  author = {Armen E. Allahverdyan and Greg Ver Steeg and Aram Galstyan},
  title = {{Memory-induced mechanism for self-sustaining activity in networks.}},
  year = {2017},
  eprint = {1712.07844},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1712.07844},
}
Figure