Source-linked AI summary
Robust dynamic classes revealed by measuring the response function of a social system
R. Crane, D. Sornette
TL;DR
The paper asks how individual waiting times and social interactions shape collective bursts of activity. It models YouTube viewing with a long-memory response kernel and an epidemic cascade, then measures relaxation across nearly 5 million videos. Most activity is Poisson-like, while hundreds of thousands of bursts follow power-law relaxation whose exponents cluster into three classes.
Problem
The study seeks to distinguish how endogenous and exogenous influences modify individual waiting-time distributions in collective social activity.
Method
The paper combines a power-law waiting-time memory kernel with an epidemic branching process modeled by a self-excited Hawkes process, applied to YouTube viewing time-series.
Results
Approximately 90% of videos are inactive or statistically Poisson, while about 500,000 others exhibit power-law herding with exponent groups near p ≈1.4, 0.6, and 0.2, consistent with θ = 0.4 ± 0.1.
Takeaways & Limitations
The clustered relaxation exponents provide a robust classification of collective human dynamics through the transformation of individual waiting times by endogenous and exogenous forces.
Abstract
from arXiv · showhide
We study the relaxation response of a social system after endogenous and exogenous bursts of activity using the time-series of daily views for nearly 5 million videos on YouTube. We find that most activity can be described accurately as a Poisson process. However, we also find hundreds of thousands of examples in which a burst of activity is followed by an ubiquitous power-law relaxation governing the timing of views. We find that these relaxation exponents cluster into three distinct classes, and allow for the classification of collective human dynamics. This is consistent with an epidemic model on a social network containing two ingredients: A power law distribution of waiting times between cause and action and an epidemic cascade of actions becoming the cause of future actions. This model is a conceptual extension of the fluctuation-dissipation theorem to social systems, and provides a unique framework for the investigation of timing in complex systems.
INTRODUCTION
The paper investigates how collective human activity reflects individual waiting times, social interactions, and external influences. It distinguishes endogenous bursts driven by cumulative internal effects from exogenous bursts responding to major external events, using YouTube viewing data as a quantitative setting.
- Motivation: Collective behavior is difficult to explain because many factors influence individuals’ decisions to act.Aggregated activity usually produces seasonal or simple patterns, but can also generate large-scale cultural and financial trends.
- Research aim: The study measures a social system’s response function to distinguish bursts caused by cumulative endogenous factors from responses to exogenous perturbations.This approach examines how interactions and external influences modify individual waiting-time distributions.
- Endogenous and exogenous bursts: Exogenous activity rises suddenly after events such as the Asian tsunami, whereas endogenous activity shows precursory growth and nearly symmetric decay, as in interest surrounding a Harry Potter movie.Both processes produce bursts, but their post-peak relaxation dynamics are expected to differ.
- Empirical setting: The dataset contains nearly 5 million YouTube activity time-series collected sub-daily over 8 months.Viewing can arise randomly, exogenously when a video is featured, or endogenously when it is shared.
THE MODEL
The model combines long-memory waiting times between individual causes and actions with an epidemic cascade through a social network. A Hawkes conditional Poisson process represents the resulting view rate and separates direct, propagated, and spontaneous contributions.
- Memory kernel: The model’s first ingredient is a power-law waiting-time distribution represented by a long-memory response function.It captures the latent impact of chance, media, links, and social influences on viewing activity.
- Memory kernel: The memory kernel φ(t) describes the waiting-time distribution between an individual’s cause and direct viewing action.The exponent θ is the theory’s key parameter and is determined empirically.
- Epidemic cascade: The second ingredient is an epidemic branching process in which one person’s attention spreads to others and becomes a trigger for future attention.The cascade is modeled as a self-excited Hawkes conditional Poisson process.
- View-rate formulation: The instantaneous view rate combines an exogenous source V(t) with influences generated by prior viewers through the memory kernel.The parameter µ_i represents the number of potential viewers directly influenced by person i after time t_i.
Predictions of the model
The model classifies activity by disturbance type—endogenous or exogenous—and network susceptibility—critical or sub-critical. These combinations predict distinct relaxation behaviors and peak-view fractions, enabling time-series to be grouped by their exponents.
- Classification: The model links four behavioral categories to disturbance type, network susceptibility, and a common memory exponent θ.The categories are endogenous or exogenous crossed with critical or sub-critical network conditions.
- Exogenous sub-critical: In the exogenous sub-critical class, mean influence ⟨µ_i⟩ is below 1, so activity remains close to the direct memory response and does not cascade far.The disturbance generates activity through only the first few generations.
- Exogenous critical: In the exogenous critical class, ⟨µ_i⟩ is close to 1, allowing propagation through many generations and producing a renormalized, slower response.The network is considered “ripe” for the video under this condition.
- Endogenous classes: Endogenous critical bursts combine word-of-mouth growth with a ripe network, producing a renormalized response before and after the activity peak.Endogenous sub-critical activity is instead expected to be dominated by fluctuations and simple stochastic behavior.
- Grouping criterion: The fraction of views on the peak day decreases from exogenous sub-critical to exogenous critical and becomes very small for endogenous critical activity.This peak fraction provides the basis for sorting time-series into three groups for exponent analysis.
RESULTS
Most videos show little or statistically random activity, while roughly 10% exhibit nontrivial herding with three robust power-law relaxation classes. Their exponents and precursory dynamics agree with the epidemic model’s predictions.
- Approximately 500,000 videos exhibit nontrivial herding behavior that follows three predicted power-law relations.These videos comprise the other roughly 10% of the dataset.
- Class 1 contains bursts with 80% ≤ F ≤ 100%, whereas Class 2 contains bursts with 20% < F < 80%.F is the fraction of views occurring on the most active day relative to the total count.
- The three relaxation-exponent groups center near p ≈ 1.4, 0.6, and 0.2 and match epidemic-model predictions using θ = 0.4 ± 0.1.The exponent clusters remain robust when the classification thresholds change.
- The endogenous-critical group shows significantly more precursory growth, while the two exogenous groups show very little.This precursory pattern is predicted by the epidemic branching model and independently tests the relaxation-based classification.
- Videos near p = 1.4 decay faster after the peak than videos near p = 0.6 or p = 0.2.The comparison follows directly from grouping videos by their post-peak relaxation exponent.
DISCUSSION
The study argues that epidemic models robustly classify collective human dynamics by combining individual waiting-time distributions with endogenous and exogenous forces. It further proposes applications to content labeling, commercial products, and marketing-response measurement.
- Collective human dynamics can be robustly classified by epidemic models as waiting-time distributions are transformed by endogenous and exogenous forces.
- The three classes support qualitative labels of viral, quality, and junk videos based on precursory growth, epidemic spread, and relaxation exponents.Viral corresponds to endogenous critical, quality to exogenous critical, and junk to exogenous sub-critical dynamics.
- The method could extend from videos to books, movies, and commercial products using sales as a proxy for activity relaxation.
- The proposed classification is robust because it uses scale-free activity dynamics rather than response magnitude or qualitative judgment.This supports identifying both mass-appeal content and content relevant to specialized communities.
- Marketing effectiveness could be quantified by measuring sales responses to exogenous marketing events.