French fake news propagation: multi-level assessment and classification

The landscape of news consumption has undergone significant transformations with the emergence of online social media platforms, leading to an exponential increase in the volume and speed of information dissemination. These platforms facilitate the rapid spread of information across vast networks with minimal quality control measures. The research on automatic Fake news detection methods improved tremendously and is still a promising application of recent machine learning methods. Those methods led to the development of Fake News Datasets, the most notable being FakeNewsNet, Twitter15 and Twitter16. However, Fake news detection methods focused primarily on English Twitter datasets. In this regard, we built a French fake news dataset comprising 15 independent and diverse events totaling over one million tweets. We then constructed the events’ Twitter propagation graph to conduct an extensive statistical analysis building upon previous findings on English datasets. Finally, we constructed the tweet cascade propagation tree and explored different fake news detection methods. The multi-level statistical analysis reveals that the propagation of Fake News is different from True News at all three levels of Network Analysis (microscopic, mesoscopic and macroscopic levels). These findings corroborate earlier observations, indicating that fake news exhibit broader and quicker dissemination patterns, often involving less verified but strategically influential users. The modeling experiments on the tweets cascades dataset showed that simple machine learning models can classify efficiently True News vs. Fake News but the paper also shows a lack of generalization when dealing with more nuanced definitions of fake news.