Bibliographic Details
| Title: |
Deep active learning for misinformation detection using geometric deep learning |
| Authors: |
Barnabò, Giorgio, Siciliano, Federico, Castillo, Carlos, Leonardi, Stefano, Nakov, Preslav, Martino, Giovanni Da San, Silvestri, Fabrizio |
| Source: |
Online Social Networks and Media; 20230101, Issue: Preprints |
| Abstract: |
Human fact-checkers currently represent a key component of any semi-automatic misinformation detection pipeline. While current state-of-the-art systems are mostly based on geometric deep-learning models, these architectures still need human-labeled data to be trained and updated - due to shifting topic distributions and adversarial attacks. Most research on automatic misinformation detection, however, neither considers time budget constraints on the number of pieces of news that can be manually fact-checked, nor tries to reduce the burden of fact-checking on - mostly pro bono - annotators and journalists. The first contribution of this work is a thorough analysis of active learning (AL) strategies applied to Graph Neural Networks (GNN) for misinformation detection. Then, based on this analysis, we propose Deep Error Sampling (DES) - a new deep active learning architecture that, when coupled with uncertainty sampling, performs equally or better than the most common AL strategies and the only existing active learning procedure specifically targeting fake news detection. Overall, our experimental results on two benchmark datasets show that all AL strategies outperform random sampling, allowing – on average – to achieve a 2% increase in AUC for the same percentage of third-party fact-checked news and to save up to 25% of labeling effort for a desired level of classification performance. As for DES, while it does not always clearly outperform other strategies, it still reduces variance in the performance between rounds, resulting in a more reliable method. To the best of our knowledge, we are the first to comprehensively study active learning in the context of misinformation detection and to show its potential to reduce the burden of third-party fact-checking without compromising classification performance. |
| Database: |
Supplemental Index |
