BERT-Based Fake News Detection: A Transformer-Driven Approach for Misinformation Classification on Twitter

Abstract

The rapid spread of fake news on social media platforms, particularly Twitter, poses a critical challenge to information credibility. This research presents an advanced fake news detection framework leveraging deep learning models, including XGBoost, CNN-RNN, BERT, and RoBERTa + GNN, to enhance detection accuracy. Our approach integrates content-based analysis, social context features, and explainable AI techniques (SHAP, LIME) for robust classification.We trained and evaluated our models on the FakeNewsNet, PolitiFact, and Kaggle fake news datasets, employing state-of-the-art feature engineering techniques such as semantic embeddings (RoBERTa, XLNet), sentiment analysis, and network propagation modeling. Experimental results demonstrate that our RoBERTa + GNN model achieves the highest accuracy of 98.7%, outperforming BERT (98.0%), CNN-RNN (84.0%), and XGBoost (81.0%). The precision, recall, and F1-scores of our models also indicate strong classification performance, with RoBERTa + GNN achieving an F1-score of 98.4%.By integrating explainability techniques, we ensure model transparency, allowing insights into the key linguistic and contextual factors influencing classification. This research contributes to improving automated misinformation detection, reducing the impact of fake news, and supporting real-time deployment for social media monitoring. Future work includes expanding cross-lingual capabilities and enhancing early detection using temporal features.