Impact of individual actions on the collective response of social systems

1. Problem Statement

The paper addresses a fundamental question in social physics: How does the activity ($A$) of an individual relate to the collective response ($R$) they trigger within a social system?
While previous studies (e.g., on Twitter) have observed that individual actions can trigger spontaneous collective reactions, the specific functional relationship and the "efficiency" of these interactions remain an open question. The authors aim to determine if the collective response is independent of individual activity or if it depends on the actor's specific features and the volume of their actions. They seek to explain the universal structure observed in the distribution of efficiency ($\eta = R/A$) across different types of social systems.

2. Methodology

Data Sources

The study analyzes three distinct social systems to ensure the findings are not domain-specific:

1. Twitter: 14 datasets covering various conversations (elections, conflicts, sports) in Spain, Turkey, Palestine, Argentina, and Colombia.
   • Activity ($A$): Number of original tweets.
   • Response ($R$): Number of retweets.
2. Scientific Citations: Data from Web of Science covering authors from 14 countries (2008–2017).
   • Activity ($A$): Number of publications.
   • Response ($R$): Total citations received.
3. Wikipedia: English Wikipedia edit history (2001–2008).
   • Activity ($A$): Total editions made by a user.
   • Response ($R$): Editions made by others on that user's personal page (treated as messages directed to the user).

The Efficiency Metric

The core metric is efficiency ($\eta$), defined as the ratio of response to activity:
$$\eta = \frac{R}{A}$$
The authors analyze the probability density function (PDF) of $\eta$, denoted as $f(\eta)$. Empirically, this distribution exhibits a universal concave shape with two distinct regimes: $\eta < 1$ (low efficiency) and $\eta > 1$ (high efficiency).

Theoretical Framework: Three Minimal Statistical Models

To explain the empirical distributions, the authors propose three domain-independent models, progressing from simple to complex assumptions regarding the dependence between $A$ and $R$:

1. Independent Variables Model (InV):

   • Assumption: $A$ and $R$ are statistically independent random variables. The system's response does not depend on the specific actor or their activity level.
   • Method: The joint probability is $p(R, A) = p(R)p(A)$. The efficiency distribution is derived analytically using random variable algebra.
   • Goal: To test if the universal shape arises simply from the ratio of two independent heavy-tailed distributions.

2. Identical Actors Model (IdA):

   • Assumption: All actors are identical, but the total response $R$ depends on the number of actions $A$. The system reacts to each single action ($r$) independently, and $R$ is the sum of $A$ partial responses ($R = \sum_{i=1}^A r_i$).
   • Method: $R$ is modeled as the $A$-fold convolution of the single-action response distribution $p(r)$. This introduces a correlation between $R$ and $A$ solely due to the summation process.
   • Goal: To test if the dependence arises simply because more actions lead to a higher aggregate response, regardless of who performs them.

3. Distinguishable Actors Model (DiA):

   • Assumption: Actors are distinguishable; their specific features influence the system's response. The response to a single action depends on the actor's features (vector $\vec{s}$).
   • Method: Applied specifically to Twitter, where the feature $\vec{s}$ is the number of followers ($F$). The response to a single action follows a Binomial distribution $B(r; F, p_{inf})$, where $p_{inf}$ is an infection probability. The aggregate response is derived by marginalizing over the distribution of followers.
   • Goal: To test if individual heterogeneity (e.g., popularity) is necessary to explain the full efficiency distribution.

Computational Techniques

The authors employed three methodologies to compute the theoretical distributions:

• Monte-Carlo (MC) Simulation: Direct implementation of model assumptions.
• Discrete Computation: Using discrete joint probability distributions and convolution.
• Analytical Derivation: Deriving closed-form expressions for the PDF of efficiency (possible for InV and IdA models under power-law assumptions).

3. Key Contributions

1. Discovery of Universal Structure: The paper confirms that the efficiency distribution ($f(\eta)$) possesses a universal structure across three fundamentally different social systems (social media, collaborative editing, and academic citations), characterized by a concave shape with distinct behaviors for $\eta < 1$ and $\eta > 1$.
2. Development of a Hierarchical Modeling Framework: The authors introduce a parsimonious framework of three models (InV, IdA, DiA) that isolate specific mechanisms (independence, aggregation, and heterogeneity) to explain social dynamics.
3. Analytical Expressions: The paper provides exact analytical expressions for the efficiency distribution under the InV and IdA models, allowing for rapid calculation and theoretical insight without relying solely on simulations.
4. Identification of Regime-Specific Drivers: The study demonstrates that no single model explains the entire distribution. Instead, different mechanisms dominate different parts of the distribution:
   • Low Efficiency ($\eta < 1$): Best explained by the InV model (independence).
   • High Efficiency ($\eta > 1$): Best explained by the IdA and DiA models (dependence on activity and actor features).

4. Results

• Scientific Citations: The InV model provided the best fit. The independence assumption held well, likely because citations are assigned to multiple authors and influenced by many external factors (quality, field, reputation), masking the direct dependence between an author's publication count and their citation count.
• Twitter and Wikipedia:
  • The InV model captured the general shape but failed to reproduce the right tail (high efficiency) accurately.
  • The IdA model significantly improved the fit for the right tail ($\eta > 1$) by accounting for the correlation between $A$ and $R$ (more actions = more response). However, it failed to fit the left tail ($\eta < 1$), predicting too few low-efficiency users.
  • The DiA model (using follower count as a proxy for influence) successfully reproduced both the left and right tails of the distribution for Twitter. It also accurately replicated the empirical Spearman's rank correlation between $A$ and $R$.
• Correlation Analysis: The discrepancy between the InV model and empirical data was found to be positively correlated with the empirical $A-R$ correlation. This confirms that ignoring the dependence between activity and response leads to errors, particularly in systems with high interaction dynamics like Twitter.
• Mechanism of Low Efficiency: The failure of the IdA model in the left tail suggests that users with low efficiency do not simply have "fewer" actions; rather, their actions have a fundamentally lower probability of triggering a response compared to high-efficiency users. This implies that actor heterogeneity (features like follower count) is crucial for explaining the full spectrum of efficiency.

5. Significance

• Theoretical Insight: The paper moves beyond simple diffusion models to establish that the relationship between individual activity and collective response is non-uniform. It is not a single law but a composite of independent stochastic processes (for low performers) and feature-dependent aggregation (for high performers).
• Methodological Utility: The proposed models serve as robust null models for social physics. Researchers can use the InV, IdA, and DiA baselines to determine if a specific domain phenomenon requires complex, domain-specific explanations or if it can be explained by minimal statistical assumptions.
• Practical Implications:
  • For Social Media Platforms: Understanding that high-efficiency users are distinct (distinguishable) suggests that algorithms targeting "viral" content must account for user features (e.g., follower count) rather than just content volume.
  • For Scientific Evaluation: The success of the InV model for citations suggests that citation counts are less dependent on the sheer volume of publication and more on independent factors, supporting the use of citations as a metric that is somewhat decoupled from simple activity levels.
• Generalizability: By demonstrating universality across Twitter, Wikipedia, and Citations, the findings suggest that the interplay between individual agency and collective reaction follows fundamental statistical laws applicable to diverse complex social systems.