People Lie, Actions Don't! Modeling Infodemic Proliferation Predictors among Social Media Users

Chahat Raj, Priyanka MeelarXiv 2021 · OpenReview

TL;DR

This paper investigates factors that predict the spread of fake news on social media, particularly during the COVID-19 pandemic. By analyzing Twitter data, the study found that sentiment polarity and user gender are significant predictors of fake news, while visual media's role is inconclusive. These findings offer practical insights for developing more effective fake news detection algorithms.

What Problem Does This Paper Address?

The proliferation of misinformation and disinformation on social media, especially during critical events like the COVID-19 pandemic, poses a significant challenge. While many studies focus on classifying fake news, there's a notable gap in understanding the underlying characteristics and factors that contribute to its spread and acceptance. This paper aims to bridge this gap by statistically identifying and analyzing key predictors that differentiate fake news from real news on platforms like Twitter, moving beyond simple classification models to a deeper understanding of the "why" behind fake news propagation.

Methods

The researchers employed a mixed-methods approach, formulating fourteen hypotheses to explore the relationships between various factors and fake news. They utilized Chi-square tests for nine qualitative hypotheses, examining categorical variables like sentiment, gender, and media usage. For five quantitative hypotheses, they employed Analysis of Means and calculated confidence intervals to assess user-specific features such as follower count, friend count, status count, retweet count, and favorite count. The study also incorporated a gender predictor algorithm and sentiment analysis tools to extract relevant features from the data.

Datasets

The study utilized two real-world Twitter datasets related to the COVID-19 pandemic:

  • MediaEval 2020: This benchmark dataset contains tweets classified into categories like 5G coronavirus conspiracy, other conspiracy, and non-conspiracy. For this study, tweets were broadly categorized into "real" (non-conspiracy) and "fake" (conspiracy-related).
  • CovidHeRA: This dataset comprises tweets related to fake remedies, preventive measures, and treatments for COVID-19, all categorized as fake news.

Key Insights and Ideas

  • Sentiment is a strong signal: The study found a significant bias in the sentiment of fake news, with neutral and negative sentiments being more prevalent than positive ones. This suggests that the emotional tone of a tweet can be a powerful indicator of its veracity.
  • Gender matters: User gender was identified as a significant predictor, with a notable bias observed in the propagation of fake news.
  • Visual media's role is murky: The presence of visual media (pictures/videos) did not consistently prove to be a reliable indicator for distinguishing fake from real news, with its impact being inconclusive across different analyses.
  • Network structure is revealing: Twitter-specific factors like follower count and retweet count showed significant differences between fake and real news. Users spreading fake news tend to have fewer followers, while the number of retweets is a key indicator of propagation.
  • Beyond classification: The research moves beyond simply identifying fake news to understanding the factors that drive its spread, offering a more nuanced perspective for future detection models.

Key Results

The study's statistical analysis led to the rejection of null hypotheses for several key factors:

  • Sentiment: A significant bias was found in the sentiment of fake news, with neutral and negative sentiments being more prevalent.
  • Gender: A significant bias was observed in the gender of users involved in fake news propagation.
  • Media Usage: A significant difference was found in the proportion of fake and real news propagation with media usage.
  • Follower Count: A significantly distinguishable bias was found in the follower count of users spreading fake news.
  • Friends Count: A significantly distinguishable bias was found in the friends count of users spreading fake news.
  • Retweet Count: A significantly distinguishable bias was found in the retweet count of tweets.

However, the contribution of status count and favorites count was found to be disputed or inconclusive, making them less reliable predictors.

Limitations and Future Work

The authors acknowledge that the study's findings for certain hypotheses, particularly those involving the interaction of gender and media usage, yielded contradictory results across the two datasets, making them inconclusive. They also note that the "number of statuses" and "number of favorites" variables were not suitable for definitive classification due to inconsistent trends.

Future work suggested by the authors includes:

  • Discovering more identifiers: Encouraging researchers to identify additional factors that can characterize false information.
  • Developing advanced algorithms: Motivating the design of algorithms that leverage the discovered dependencies and combined decisions of these factors for more effective fake news detection.
  • Expanding the scope: Exploring these factors across a wider range of social media platforms and contexts to build a more comprehensive understanding of infodemic proliferation.

Summary updated

Abstract

Social media is interactive, and interaction brings misinformation. With the growing amount of user-generated data, fake news on online platforms has become much frequent since the arrival of social networks. Now and then, an event occurs and becomes the topic of discussion, generating and propagating false information. Existing literature studying fake news primarily elaborates on fake news classification models. Approaches exploring fake news characteristics and ways to distinguish it from real news are minimal. Not many researches have focused on statistical testing and generating new factor discoveries. This study assumes fourteen hypotheses to identify factors exhibiting a relationship with fake news. We perform the experiments on two real-world COVID-19 datasets using qualitative and quantitative testing methods. This study concludes that sentiment polarity and gender can significantly identify fake news. Dependence on the presence of visual media is, however, inconclusive. Additionally, Twitter-specific factors like followers count, friends count, and retweet count significantly differ in fake and real news. Though, the contribution of status count and favorites count is disputed. This study identifies practical factors to be conjunctly utilized in the development of fake news detection algorithms.
BibTeX
@misc{raj2021people,
  author = {Chahat Raj and Priyanka Meel},
  title = {{People Lie, Actions Don't! Modeling Infodemic Proliferation Predictors among Social Media Users}},
  year = {2021},
  eprint = {2111.11955},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.11955},
}
Figure