TL;DR
This paper proposes a unified mathematical model that accurately describes information spreading on social media, driven by a novel interplay of social reinforcement and weakening effects. This model resolves long-standing controversies in the field and reveals how the clustered nature of social networks enables rapid, diverse information dissemination.
What Problem Does This Paper Address?
Understanding how information spreads on social media is crucial, given its pervasive influence on society. Previous models, often drawing analogies to disease spread (like the SIR model), struggled to capture the nuances of human behavior in online environments. Specifically, the "social reinforcement effect" – where repeated exposure increases the likelihood of sharing – was a prominent theory. However, empirical observations showed this wasn't the whole story. Contradictory findings emerged: some studies suggested reinforcement, while others indicated social weakening (e.g., reduced sharing with more common neighbors) or that repeated exposure could decrease sharing probability. This paper aims to reconcile these conflicting observations and provide a more comprehensive, universally applicable model for information diffusion on social networks.
Methods
The authors propose a new mathematical equation for information spreading, represented by the retweeting probability $\beta_i(x)$ of message $i$ at the $x$-th exposure:
$\beta_i(x) = \alpha_i x (1 - \gamma)^{f_i(x)}$
Here's a breakdown of the terms:
- $\alpha_i$: Represents the intrinsic spreading power of message $i$.
- $x$: The number of exposures to the message. This term captures a basic linear reinforcement effect.
- $(1 - \gamma)$: Represents the proportion of a user's friends who are not yet aware of the message, thus forming potential new audiences. $\gamma$ is the average proportion of common neighbors between any two users.
- $f_i(x)$: A "user sensitivity" term that accounts for users' uncertainty in estimating the gains of retweeting. The authors found that $f_i(x) = x^{\omega_i}$ (where $\omega_i$ is a parameter) best fits their empirical data, leading to the final form: $\beta_i(x) = \alpha_i x (1 - \gamma)^{\omega_i}$.
This equation is derived by considering information spreading as the propagation of human behaviors, where gains from mimicking peers are central. The authors also develop a method to determine the form of $f_i(x)$ by combining theoretical boundary conditions with empirical data, leading to a specific form that captures the observed non-monotonicity of retweeting probability (rising then falling).
Datasets
The study utilizes data from three major social media platforms:
- WeChat: Large-scale spreading, 2019.06-2019.10.
- Weibo: Large-scale spreading, 2017.01-2017.10.
- Twitter: Small-scale spreading, 2010.10.
The datasets are described as publicly accessible and strictly de-identified. A total of 7.45 billion users are cumulatively encompassed by these datasets. For analysis, trajectories with at least 5 data points and involving at least 100 retweeting users were selected.
Key Insights and Ideas
- Unified Mechanism: The core insight is that information spreading is driven by a universal interplay of social reinforcement (more exposure, more likely to share) and social weakening (less novelty, less incentive to share). This resolves the paradox of conflicting previous findings.
- Non-Monotonic Spreading: The proposed model predicts and empirically confirms that retweeting probability typically rises and then falls with the number of exposures. This is a crucial departure from simpler models.
- Weak Tie Effect: The model naturally incorporates and explains the "weak tie effect," where connections with fewer common neighbors (lower $\gamma$) facilitate broader information spread.
- Network Structure Matters: The clustered nature of social networks is shown to lead to rapid, high-frequency bursts of information, but with relatively small coverage per burst. This is key to social media's capacity for diverse information dissemination.
- Behavioral Focus: The paper emphasizes that understanding information spread requires focusing on human behavior and the "gains of mimicking peer behaviors" rather than just the information itself.
Key Results
- The proposed equation $\beta_i(x) = \alpha_i x (1 - \gamma)^{\omega_i}$ accurately describes the empirical spreading trajectories of messages across WeChat, Weibo, and Twitter.
- The model successfully explains phenomena that were controversial or unexplained by previous theories, such as the social reinforcement and social weakening effects, and the role of common neighbors.
- Simulations based on the model demonstrate that information outbreaks are relatively small in size but frequent, leading to high diversity and capacity for information dissemination, a stark contrast to the large, infrequent outbreaks predicted by social reinforcement models.
- The model's prediction for the turning point of the retweeting probability curve aligns well with empirical data.
Limitations and Future Work
The authors acknowledge several limitations:
- New Comments/Quote Tweets: The model does not explicitly account for how users adding new comments when forwarding messages (e.g., Twitter's quote tweet feature) might influence dissemination.
- Bots and Manipulated Content: The impact of artificially manipulated bots on social media and their influence on spreading dynamics is not fully addressed, though the authors suggest their model could be extended to explore this.
- Integration with AI: The paper suggests that integrating the proposed model with artificial intelligence algorithms could lead to better prediction of information spreading size, which remains an important area for future research.
They also note that while the model fits a vast amount of data, further confirmation through analysis of more comprehensive datasets would be beneficial to establish its universality.
Summary updated