Comparative analysis of machine learning methods for automated fake news detection

The spread of fake news on the internet undermines public trust, polarizes opinion, and influences critical decisions. This creates an urgent need for automated detection. This paper presents a comparative analysis of traditional machine learning (ML) and deep learning (DL) methods for fake news detection. We evaluated five models: Support Vector Machine (SVM), Random Forest, a GRU-based Recurrent Neural Network (RNN), a Long Short-Term Memory (LSTM) network, and an LSTM autoencoder. We used a structured hyperparameter search protocol with deterministic train, validation, and test splits whose indices are saved and reused across models. Hyperparameter optimization is performed, deep models select both hyperparameter combinations and epoch count from validation performance with early stopping, and final results are reported only once on a held-out test split. In the merged Kaggle Fake News and ISOT dataset, the best saved test result is achieved by the GRU-based RNN, with 98.08% accuracy and a 98.02% F1-score, followed by SVM at 96.23% accuracy and a 96.08% F1-score. On the Combined_Corpus dataset, the GRU-based RNN also achieves the strongest saved result with 98.45% accuracy and a 98.52% F1-score, with the LSTM close behind at 98.28% accuracy and a 98.36% F1-score. Overall, the results show that recurrent models are highly effective under this evaluation protocol, with the GRU-based RNN delivering the strongest saved results across both datasets, while the relative ordering of the remaining models still depends on the corpus and preprocessing pipeline.