Surges of collective human activity emerge from simple pairwise correlations

Christopher W. Lynn, Lia Papadopoulos, Daniel D. Lee, Danielle S. BassettarXiv 2018 · arXiv

Introduction

Human activity is characterized by "bursts" and "surges"—moments where collective participation far exceeds average rates. While these spikes are often attributed to external schedules (like the work week) or specific events, this research explores whether they can emerge organically from internal dynamics. By applying the Ising model from statistical physics to social data, the authors demonstrate that simple pairwise correlations between individuals account for the vast majority of collective behavior.

The Failure of Independence

Traditional models often assume individuals act independently. However, such models fail spectacularly at predicting large-scale surges. For instance, in an email dataset of 100 scientists, an independent model predicts a 7-person surge would occur once every $10^{20}$ seconds—effectively never. In reality, these events occur frequently, driven by the underlying social fabric.

The pairwise maximum entropy model accurately predicts the probability of activity surges (a) and shows a strong relationship between observed correlations and learned interaction parameters (b).

Methods: The Maximum Entropy Approach

The researchers used the Principle of Maximum Entropy to construct a model that captures observed pairwise correlations while remaining as unbiased as possible regarding higher-order interactions. This results in a model with two primary parameters:

  1. External fields ($h_i$): An individual's intrinsic bias toward activity.
  2. Pairwise interactions ($J_{ij}$): The functional influence individuals exert on one another.

The model was tested across four diverse datasets: email activity, private messaging among students, face-to-face contacts at a conference, and music streaming on Last.fm.

Key Results: Pairwise is Plenty

The study found that pairwise interactions are sufficient to explain the bulk of collective correlation. In groups of 10, the pairwise model captured approximately 89% of the "multi-information" (a measure of total network correlation). Even when scaled to 100 individuals, the model accounted for roughly 75% of the correlation.

Model parameters and performance across datasets. (a) Visualization of learned interactions. (c-d) The pairwise model (red) captures a significantly higher fraction of multi-information than models based solely on external schedules (green). (e) Scaling of information capture with group size.

Interestingly, the importance of external "schedules" varies by context. In email, weekly rhythms explain about 67% of correlations. In contrast, for private messaging, schedules explain only 5%, indicating that conversational bursts are almost entirely driven by direct person-to-person interactions. This suggests that the pairwise model is not just a statistical fit, but a reflection of real-world social influence.

Conclusion

This work suggests that complex, large-scale social phenomena can be understood through the lens of simple, fine-scale interactions. While the model currently assumes synchronous activity and constant social rules, it provides a powerful, parsimonious framework for predicting collective human behavior across various social contexts, outperforming complex schedule-based models with significantly fewer parameters.

Summary updated

BibTeX
@misc{lynn2018surges,
  author = {Christopher W. Lynn and Lia Papadopoulos and Daniel D. Lee and Danielle S. Bassett},
  title = {{Surges of collective human activity emerge from simple pairwise correlations}},
  year = {2018},
  eprint = {1803.00118},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.00118},
}
Figure