🤖 machine learning

Predicting, Evaluating, and Explaining Top Misinformation Spreaders via Archetypal User Behavior

This paper proposes a framework for predicting, evaluating, and explaining top misinformation spreaders by modeling three distinct user archetypes (amplifiers, super-spreaders, and coordinated accounts), demonstrating that super-spreader traits dominate top rankings while temporal dynamics and explainable AI enhance predictive accuracy and interpretability for proactive content moderation.

Original authors: Enrico Verdolotti, Luca Luceri, Silvia Giordano

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Enrico Verdolotti, Luca Luceri, Silvia Giordano

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

On social media, information travels at the speed of a thought, but not everyone who shares it plays the same role. Some people create the original stories, while others simply pass them along, and a few work together to make sure a message reaches as many eyes as possible. When false information spreads, it is rarely the fault of a single person; instead, it is the result of a complex web of behaviors where different types of users amplify, coordinate, and launch narratives. Understanding who these people are and how they operate is crucial because false stories can distort public opinion, erode trust in science, and cause real-world harm. The challenge for platforms is that by the time a false story is flagged and removed, it has often already traveled too far to be stopped. To stop the spread before it gains momentum, researchers need a way to identify the specific accounts most likely to cause damage, not by guessing their intentions, but by watching how they actually behave.

A team of researchers set out to solve this problem by looking at the digital footprints left behind by users who spread low-credibility content. They focused on three distinct patterns of behavior, or "archetypes," that these users tend to follow. The first group, which they call amplifiers, rarely creates their own posts. Instead, they act as a megaphone, constantly resharing content made by others to boost its reach. The second group, known as super-spreaders, are the original creators of viral content. They write posts that consistently get shared thousands of times, acting as the spark that ignites a fire. The third group consists of coordinated accounts, which are groups of users working in sync to push the same message at the same time, creating an illusion of widespread public agreement. The researchers realized that these roles are not always separate; a single user might act as a super-spreader for one story and an amplifier for another, or a group might coordinate their actions to mimic a single influential voice.

To test how well they could predict who would spread misinformation next, the researchers analyzed two massive collections of data from Twitter during the pandemic. One dataset contained conversations from Italy, while the other was a global collection of posts in many languages. They first labeled the content based on the trustworthiness of the websites linked in the posts, using a system that rates news domains from zero to one hundred. They then built different mathematical models to rank users based on the three behavioral archetypes. Some models counted how many times a user reshared false stories, while others looked at how early a user shared a story after it was posted, or how closely a user's activity matched that of others in a coordinated group. They also created a new, time-sensitive version of a classic influence score that accounts for how a user's activity changes over days and weeks, rather than just looking at a single snapshot in time.

The results showed that the most effective way to find the biggest troublemakers was to look for the super-spreaders. When the researchers ranked users based on their ability to generate original content that others would reshare, these accounts consistently appeared at the very top of the list. These are the individuals responsible for the initial burst of virality that allows false narratives to take hold. However, as the researchers looked further down the list of influential users, the picture became more complex. In the middle and lower ranks, the behavior of amplifiers and coordinated groups became much more prominent. This suggests that while a few super-spreaders start the fire, it is the amplifiers who keep it burning and the coordinated groups who fan the flames across the network.

To get a clearer picture of how these different behaviors work together, the team trained a machine learning model that combined all three archetypes into a single system. They then used a technique that explains how the model makes its decisions, allowing them to see which behaviors mattered most. They found that the model relied heavily on the number of times a user reshared low-credibility content, a trait associated with amplifiers. This was surprising because one might expect the creators of viral content to be the most important signal. Instead, the data showed that the act of resharing is a powerful indicator of risk. The model also confirmed that the super-spreaders, identified by their ability to generate consistent engagement, were critical, especially among the very top-ranked users. As the ranking moved down, the influence of coordinated behavior and the frequency of resharing grew stronger, revealing a layered ecosystem where different types of actors play different roles in the same problem.

The study also tested how well these methods work when there is very little data to go on, simulating a scenario where a platform needs to make a decision quickly after a new story appears. They found that the time-sensitive influence score remained stable and accurate even with limited information, outperforming older methods that did not account for the timing of user activity. The machine learning model, however, struggled when trained on very small amounts of data, suggesting that while complex models are powerful, they need enough history to learn the patterns. The researchers concluded that the best approach for platforms is likely a combination of these tools: using simple, time-aware scores to quickly identify the most dangerous accounts, and using more detailed models to understand the full range of behaviors that sustain the spread of misinformation. By focusing on these observable patterns rather than trying to guess a user's intent, platforms can intervene earlier, targeting the specific accounts that drive the most harm before false information becomes impossible to contain.

Technical Summary: Predicting, Evaluating, and Explaining Top Misinformation Spreaders via Archetypal User Behavior

Problem Statement

The rapid dissemination of misinformation on social networks poses a significant threat to societal stability, public health, and democratic institutions. Current moderation strategies are predominantly reactive, intervening only after content has already achieved widespread circulation. This lag allows false narratives to gain legitimacy and exploit algorithmic amplification before they can be effectively countered. The core challenge addressed by this paper is the inability to proactively identify and rank users who are most likely to drive the spread of low-credibility content based on their behavioral footprints, rather than waiting for specific content to be flagged. While previous research has focused heavily on "super-spreaders" (content creators) and "coordinated accounts," the role of "amplifiers" (users who primarily reshare) remains underexplored and loosely defined.

Methodology

1. Behavioral Archetypes

The authors formalize three distinct, non-mutually exclusive user archetypes responsible for misinformation diffusion:

Amplifiers: Users who generate little original content but significantly boost reach through extensive resharing.
Super-spreaders: Highly influential users whose original content consistently achieves viral reach through high engagement and resharing.
Coordinated Accounts: Groups of users acting in collaboration to artificially amplify specific narratives, exhibiting synchronized posting and behavioral similarity.

2. Data and Credibility Assessment

The study utilizes two datasets collected from Twitter during the COVID-19 pandemic:

VaccinItaly: Italian-language tweets (Dec 2020 – Oct 2021).
COVID-19 Multilanguage: Global tweets with a US bias (Nov 2020 – Nov 2021).
Credibility is assessed using NewsGuard scores assigned to URLs within posts. Posts linking to domains with scores ≤39 are classified as low-credibility. The analysis focuses exclusively on original posts and reshares, excluding replies and self-reshares.

3. Ranking Methodologies

The paper develops and compares several ranking approaches:

A. Archetype-Specific Heuristics:

Amplifiers:
- Repost Count: Total number of low-credibility reshares.
- Early Reposter Index (EaR-Index): Weights users higher if they reshare content early in the diffusion cascade.
Coordinated Accounts:
- Coordination-Centrality: Eigenvector centrality in a co-reshare similarity network.
- Coordination-Edge Weight: Maximum edge weight in the similarity network.
Super-spreaders:
- Influence Score: Total reshares received by a user's low-credibility posts.
- H-Index: The largest number $h$ such that the user has $h$ posts with at least $h$ reshares.
- Proposed Time-Aware Variants: To address the static nature of standard metrics, the authors introduce Time-Aware Influence Score (TAI-Score) and Time-Aware Social H-Index (TASH-Index). These use an exponential moving average over non-overlapping time slots to balance historical influence with recent activity.

B. Machine Learning Framework:
A regression-based "Learning to Rank" model (specifically Random Forest) is trained to predict a user's future contribution to misinformation. The input features are the scores derived from the archetype-specific methods listed above. The target variable is the total number of reshares involving low-credibility content a user engages with during the test period.

4. Evaluation

Performance is evaluated using a network dismantling approach on a low-credibility reshare network.

Quality@k: Measures the proportion of misinformation volume removed by blocking the top- $k$ ranked users.
nDCG@k: Normalized Discounted Cumulative Gain, measuring the consistency between predicted and actual user influence.
The evaluation uses a temporal split (80% training, 20% testing) to prevent data leakage.

5. Explainability

The study employs SHAP (SHapley Additive exPlanations) to interpret the machine learning model, quantifying the contribution of each behavioral feature to the final ranking predictions.

Key Results

Super-Spreaders Dominate Top Ranks: The TASH-Index (Time-Aware Social H-Index) and the Machine Learning (ML) model achieve performance comparable to the optimal dismantling strategy, particularly for the top-ranked users (small $k$ ). This confirms that a small number of content creators (super-spreaders) are responsible for a disproportionate share of misinformation dissemination.
Role of Amplifiers in Lower Ranks: As the ranking moves down (e.g., $k=1000$ ), the TAI-Score and Repost Count become increasingly significant. The ML model reveals that while super-spreaders initiate virality, amplifiers play a critical role in sustaining and extending the reach of misinformation.
Superiority of Time-Aware Metrics: The TASH-Index consistently outperforms the static H-Index and Influence Score, demonstrating that temporal dynamics are critical for accurate prediction.
Data Scarcity Robustness: The TASH-Index remains stable even with limited observation windows (as low as 30 days), whereas retraining ML models from scratch on small datasets leads to performance degradation. However, a pre-trained ML model maintains high stability across data reduction ranges.
Feature Importance (SHAP Analysis):
- Repost Count (Amplifier trait) is the single most influential feature in the ML model, highlighting the critical role of resharing behavior.
- TASH-Index and TAI-Score (Super-spreader traits) are the next most significant features.
- Coordination-Centrality gains importance in the mid-to-lower ranks, suggesting that coordinated behavior is a key driver for users who are not top-tier super-spreaders.

Significance and Claims

The paper claims three primary contributions to the field of misinformation research:

Formalization of Archetypes: It provides a rigorous definition and operationalization of three behavioral archetypes (Amplifiers, Super-spreaders, Coordinated Accounts), addressing a gap where "amplifiers" were previously loosely defined and often overlooked in favor of content creators.
Temporal Dynamics and TASH-Index: It introduces the TASH-Index, a computationally efficient, time-aware ranking method that captures the temporal evolution of user influence. The authors demonstrate that this simple heuristic performs competitively with complex machine learning models while requiring fewer resources and offering greater interpretability.
Explainable AI for Intervention: By integrating multiple archetypal features into a unified ML model and applying SHAP, the study offers a nuanced understanding of how different behaviors drive misinformation. It argues that effective moderation strategies should be tiered: targeting super-spreaders with preventive scrutiny and addressing amplifiers/coordinated accounts through systemic measures like limiting amplification or applying friction.

The authors conclude that their approach provides actionable tools for platforms to identify high-risk users proactively. They emphasize that their framework focuses on observable behavioral patterns rather than user intent, aligning with governance principles that prioritize mitigating systemic harm. The work suggests that combining simple, interpretable behavioral metrics with machine learning offers a robust path toward scalable and transparent content moderation.