Scalable and reliable anomaly detection on Dynamic Graph based on link prediction to identify disinformation
Résumé
As social networks take an ever-prominent role in information access, combating disinformation becomes increasingly important. Given current volumes of data, automated approaches for the detection of disinformation are the only ones able to offer an online solution. We formalize disinformation as an anomaly in the organic evolution of the network, in terms of users, content or coordination. To identify it, we propose a solution based on Temporal Graph Networks (TGN) adapted to the detection of anomalies and completed with a reliability module which permits users to trade precision for recall. Inheriting the performances of TGN, this solution is able to scale up, work on continuous time settings, and handle multimodality (text, image and video). Moreover, when compared to existing models, our approach outperforms state-of-the-art solutions for anomaly detection in dynamic graphs on several classical datasets. Lastly, it has been tested on a Twitter dataset of the French presidential election of 2022, providing useful insights on manipulation during the campaign.
Origine : Fichiers produits par l'(les) auteur(s)