Deep Learning-Based Fake News Detection in Social Media

Abstract

The proliferation of fake news on social media has introduced a major problem in society that is troubling public trust, political stability, health awareness and the reliability of information. Machine learning methods are inadequate in many aspects in the field of fake news classification, especially in modelling the long-range dependence and semantic meaning of text. Traditional machine learning algorithms are generally not suitable for modelling the long-range dependence and semantics of text, which is a limitation in the context of fake news classification. To overcome these drawbacks, a hybrid deep learning architecture using Bidirectional Encoder Representations from Transformers (BERT), Bidirectional Long Short-Term Memory (BiLSTM) and Attention mechanism is proposed to ensure accurate and explainable identification of fake news. The proposed architecture is designed to leverage the contextual embedding generation, sequential dependency learning, and attention-based feature prioritization for boosting the classification performance. The experimental evaluation was performed on the two publicly available FakeNewsNet and ISOT Fake News datasets, which yielded a total of 38,994 news samples. The text quality was improved through several preprocessing techniques such as lowercasing, tokenisation, lemmatisation, stop word elimination, and noise filtering. The suggested model was verified using baseline models such as Support Vector Machine (SVM) and Logistic Regression. The experimental findings demonstrated that the suggested framework performs exceptionally well, with a 99.29% accuracy rate, a 99.74% precision rate, a 98.97% recall rate, a 99.35% F1-score, and a 99.33% ROC-AUC score. The robustness and generalization capability of the model were further confirmed by cross validation, ablation analysis, evaluation of confusion matrix, SHAP explainability and McNemar statistical testing. The results show that the embedding of transformer-based contextual learning and sequential deep learning architectures strongly improves the performance of fake news detection in social media.