Detecção precoce de rumor: avaliação da eficácia do uso de redes neurais de grafo

Abstract

The field of research focused on combating misinformation on social media has incre asingly demanded robust methods for detecting rumors and fake news. In this context, this study evaluates the use of Graph Neural Networks (GNNs) for rumor detection, comparing them with traditional models from the literature, with an emphasis on early detection performance.The proposed method includes the development of a standardized module for quick algorithm testing and comparison, ensuring chronological data splits to prevent information leakage and better reflect rumor propagation in real-world scena rios. This module enables retraining and incremental inference to account for updates in the temporal features of the training data. It also stores key classification metrics, pro viding metadata about the effectiveness in early detection.Using the PHEME dataset, a heterogeneous graph was generated to model interactions between tweets and users. Two main experiments were conducted to test two GNNs, along with the LightGBM, Random Forest, and LSTM models.The results showed that GNNs achieved a better balance between precision and recall than models that process tabular data, demonstra ting their potential for real-time misinformation detection by effectively capturing the complex relationships between tweets and users. This research contributes to the field of rumor detection through graph-based learning, providing more suitable techniques for designing new algorithms by standardizing procedures focused on early detection.