Social Media Fake News Detection Using Transformer Model

Fake news's extensive proliferation on social media has become a major worldwide concern, undermining public trust in digital information and influencing elections, social movements, and health-related decisions. With the increasing reliance on online platforms for news consumption, studies reveal that false information spreads substantially faster-up to six times more quickly-than factual news on platforms like Twitter. This rapid propagation intensifies the need for robust, real-time misinformation detection mechanisms. Conventional methods for machine learning, like Logistic Regression and Random Forest has been used in this challenge but often struggle to understand the deeper contextual meanings embedded in short or ambiguous headlines. These models typically rely on surface-level features and are limited in capturing semantic nuances. The research offers a BERT (Bidirectional Encoder Representations from Transformers)based method for detecting bogus news, a cutting-edge large language model (LLM) designed for deep Natural Language Processing (NLP) applications, for the purpose of overcoming these limitations. Leveraging BERT's bidirectional encoding capabilities, the model captures complex linguistic patterns and contextual relationships, offering a more accurate classification of news headlines. In this work, a BERT-base-uncased model is finetuned using a labeled dataset of real and fake news headlines. After applying preprocessing techniques like normalization and subword tokenization, the dataset is strategically split into training, validation, and test sets. Model will be processed or trained over 5 epochs using the AdamW optimizer. The learning rate is found to be $1 \mathrm{e}-5$. Accuracy, precision, recall, F1-score, and confusion matrix are used to assess performance. Accuracy, precision, recall, $\mathbf{F 1}$-score, and confusion matrix are used to assess performance. The model achieves $\mathbf{8 8. 0 8 \%}$ accuracy and an F1- score of 88.31 %, clearly outperforming traditional models and demonstrating clear potential for integration into social media moderation tools and automated news validation systems.