Source-linked AI summary
Users Polarization on Facebook and Youtube
Alessandro Bessi, Fabiana Zollo, Michela Del Vicario, Michelangelo Puliga, Antonio Scala, Guido Caldarelli, Brian Uzzi, Walter Quattrociocchi
TL;DR
The paper asks whether different content-promotion mechanisms on Facebook and YouTube produce different patterns when users consume the same scientific and conspiracy-like videos. Using quantitative analysis of users’ interactions and consumption across both platforms, it finds similar polarization dynamics and echo-chamber formation, with behavioral patterns supporting cross-platform prediction. The study’s evidence is limited to public data and selected scientific and conspiracy-like pages.
Problem
The paper examines whether different mechanisms regulating content promotion on Facebook and YouTube lead users consuming the same contents to form homogeneous echo chambers.
Method
The study quantitatively compares consumption and interaction patterns for videos from scientific and conspiracy-like pages across Facebook and YouTube.
Results
Users form polarized, homogeneous communities with similar behavioral patterns on Facebook and YouTube, enabling models trained on one platform to predict polarization on the other.
Takeaways & Limitations
Conflicting narratives aggregate users into homogeneous echo chambers irrespective of the platform and its content-promotion algorithm.
Takeaways & Limitations
The analysis is restricted to public data and selected scientific and conspiracy-like pages, excluding users with privacy restrictions.
Abstract
from arXiv · showhide
On social media algorithms for content promotion, accounting for users preferences, might limit the exposure to unsolicited contents. In this work, we study how the same contents (videos) are consumed on different platforms -- i.e. Facebook and YouTube -- over a sample of $12M$ of users. Our findings show that the same content lead to the formation of echo chambers, irrespective of the online social network and thus of the algorithm for content promotion. Finally, we show that the users' commenting patterns are accurate early predictors for the formation of echo-chambers.
Introduction
This study examines whether different content-promotion mechanisms on Facebook and YouTube produce different patterns of polarization and echo-chamber formation. It also investigates behavioral patterns that may predict users’ polarization.
- Polarized communities can reduce viewpoint heterogeneity, an important component of strong democratic societies.
- The study compares consumption of the same videos across Facebook and YouTube to assess whether platform-specific promotion mechanisms produce homogeneous echo chambers.The videos come from scientific and conspiracy-like Facebook pages.
- The vast majority of users who initially switch between competing narratives later consume mainly one type of information and become polarized toward one narrative.
- A multinomial logistic model predicts whether users will polarize toward a narrative or continue switching between competing information.
- Conflicting narratives can lead users to aggregate into separate echo chambers despite differences between Facebook and YouTube content-promotion algorithms.
Results
Across Facebook and YouTube, users consuming conflicting narratives show similar consumption and polarization patterns, with most users forming separated echo chambers. Early commenting behavior also predicts users’ eventual polarization across platforms.
- Content consumption: 0.987: Science and Conspiracy users’ action-correlation matrices show a statistically significant, high, positive Mantel correlation.The simulated p-value is < 0.01, based on 104 Monte Carlo replicates.
- Content consumption: Users’ consumption patterns for conflicting narratives show meaningful similarities across Facebook and YouTube.The analysis compares comments and likes for the same videos across both platforms.
- Polarized and Homogeneous Communities: 93.6% of Facebook users and 87.8% of YouTube users are polarized toward one of the two conflicting narratives.Polarized users have ρ < 0.05 or ρ > 0.95; the corresponding bimodality coefficients are 0.964 and 0.928.
- Polarized and Homogeneous Communities: Users supporting conflicting narratives behave similarly in polarized communities, irrespective of platform and content-promotion algorithm.The findings describe two well-separated communities organized around competing narratives.
- Polarization dynamics: After an initial switching phase, most users who move between narratives mainly consume one information type and become polarized.Some users interact with one content type from the beginning, while others switch before preferentially attaching to one narrative.
- Prediction of Users Polarization: n = 50 provides accuracy greater than 0.80 for each class on both YouTube and Facebook, while performance increases with n.The model uses commenting activity to forecast whether users become polarized toward a narrative or continue switching between competing narratives.
- Prediction of Users Polarization: Early mobility in commenting accurately predicts users’ preferential attachment to a specific echo chamber.Monte Carlo validations reported in Table 1 confirm the model’s performance.
Discussion
The paper argues that polarization and echo-chamber formation persist across Facebook and YouTube despite different content-promotion mechanisms. It also shows that similar behavioral patterns support cross-platform prediction of users’ polarization.
- Conflicting narratives lead users to aggregate into homogeneous echo chambers irrespective of the platform or content-promotion algorithm.The study frames this as a challenge to explanations centered primarily on algorithmic personalization.
- Users supporting competing scientific and conspiracy-like narratives form polarized communities that behave similarly on Facebook and YouTube.The comparison uses the same contents across the two online social networks.
- A statistical learning model predicts whether users become polarized toward a narrative or continue switching between competing contents.The classification task distinguishes users polarized in Conspiracy, not polarized, and polarized in Science.
Methods
The study combines public Facebook and YouTube data with operational measures of user polarization, distributional analysis, and multinomial classification. It compares narrative engagement across platforms and evaluates whether early commenting behavior predicts users’ later polarization.
- Data Collection: The Facebook dataset contains 413 US public pages categorized as Conspiracy or Science, with posts and user interactions collected from January 2010 through December 2014.
- Data Collection: The YouTube dataset contains about 17K videos linked by Facebook posts supporting Science or Conspiracy narratives.
- Measures: User polarization ρu is defined as the fraction of a user’s comments on posts or videos supporting conspiracy-like narratives.Users with ρu > 0.95 are classified as polarized towards Conspiracy, while users with ρu < 0.05 are classified as polarized towards Science.
- Distributional Analysis: Bimodality is evaluated against BCcrit = 5/9 ≈ 0.555, with higher coefficients indicating bimodality and lower coefficients indicating unimodality.
- Prediction and Evaluation: A multinomial logistic model predicts probabilities for multiple categorical outcomes, while precision, recall, and accuracy evaluate classification performance using true- and false-classification counts.The model contrasts J − 1 categories with baseline category J; accuracy is computed from TP_i, TN_i, FP_i, and FN_i.
- Distributional Analysis: The study assesses distributional similarity by estimating power-law scaling exponents with maximum likelihood and comparing their posterior difference using an Empirical Bayes procedure.Posterior distributions are obtained with Metropolis-Hastings MCMC, using 50,000 iterations with 5,000 burned.