CredBERT: Credibility-aware BERT model for fake news detection

Abstract

The spread of fake news on social media poses significant challenges, especially in distinguishing credible sources from unreliable ones. Existing methods primarily rely on text analysis, often neglecting user credibility, a key factor in enhancing detection accuracy. To address this, we propose CredBERT, a framework that combines credibility scores derived from user interactions and domain expertise with BERT-based text embeddings. CredBERT employs a multi-classifier ensemble, integrating Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), BiLSTM, Logistic Regression, and k-Nearest Neighbors, with predictions aggregated using majority voting, ensuring robust performance across both balanced and imbalanced class datasets. This approach effectively merges user credibility with content-based features, improving prediction accuracy and reducing biases. Compared to state-of-the art baselines FakeBERT and BiLSTM, CredBERT achieves 6.45% and 4.21% higher accuracy, respectively. By evaluating user credibility and content features, our model not only enhances fake news detection but also contributes to mitigating misinformation by identifying unreliable sources.