Automated Fake News Detection Using TF-IDF Features and Neural Networks: A Comparative Study

Abstract

The proliferation of misinformation across digital platforms has emerged as a critical threat to democratic processes, public health decisions, and social cohesion, with fake news spreading six times faster than authentic information on social media networks. The exponential growth of unverified content necessitates automated solutions capable of real-time authenticity verification, as traditional manual fact-checking mechanisms cannot scale to address the massive volume of daily content generation. This study presents a comprehensive framework for automated fake news detection through systematic comparison of traditional machine learning algorithms and state-of-the-art deep learning architectures, implementing five machine learning models including Logistic Regression, Random Forest, Naive Bayes, K-Nearest Neighbors, and Gradient Boosting alongside four deep learning approaches featuring LSTM, Bidirectional LSTM, GRU, and CNN architectures. Our methodology incorporates TF-IDF vectorization for feature extraction and leverages two prominent benchmark datasets: the LIAR dataset containing <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 2, 8 3 6}$</tex> political statements and the ISOT dataset comprising <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 4, 8 9 8}$</tex> news articles, creating a merged dataset of 57,734 instances with balanced class distribution. Experimental results demonstrate that deep learning models significantly outperform traditional approaches, with Bidirectional LSTM achieving <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 4. 2 \%}$</tex> accuracy on the combined dataset, while ensemble methods combining multiple model predictions further enhance classification robustness by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3. 8 \%}$</tex>. The study reveals distinct linguistic patterns between fake and real news, including higher frequencies of sensational terms and emotional manipulation in misinformation content, providing valuable insights for developing more sophisticated detection mechanisms. Our comprehensive evaluation framework establishes benchmarks for cross-domain fake news detection and demonstrates the superiority of neural network architectures in capturing complex textual patterns indicative of news authenticity. These findings contribute to the development of scalable automated fact-checking systems essential for maintaining information integrity in the digital age.