Comparative Analysis of Machine Learning Algorithms for Fake News Detection on Social Media: A Systematic Literature Review

Abstract

The rapid proliferation of fake news across social media platforms constitutes one of the most severe epistemic and societal threats of the digital age. Automated detection using machine learning (ML) and deep learning (DL) techniques has become essential due to the sheer volume, velocity, and viral nature of online misinformation. This systematic literature review (SLR) synthesises 44 peer-reviewed empirical and review studies published between 2020 and 2026, following PRISMA 2020 guidelines, to evaluate the state of ML/DL-based fake news detection on social media. The corpus spans classical ML classifiers (SVM, Random Forest, Naïve Bayes, XGBoost), deep learning architectures (CNN, LSTM, BiLSTM, attention mechanisms), transformerbased models (BERT, CT-BERT, RoBERTa, FakeBERT, ABERT, AraBERT), and advanced ensemble, graph-based, and LLM-augmented approaches that integrate social context, user profiling, knowledge graphs, and contrastive learning. Findings strongly support both tested hypotheses: H₂ – deep learning and transformer models consistently outperform classical ML, with transformer ensembles achieving F1-scores of 93–99.1% across benchmark datasets including LIAR, FakeNewsNet, ISOT, Weibo, and COVID-19 corpora; and H₃ – ensemble and hybrid methods incorporating social and contextual features deliver superior robustness, with the best hybrid systems (TAM-ATOA, MG-CL, KeepUp) outperforming single-model baselines by 3–12 percentage points. Key challenges identified include dataset bias, class imbalance, multilingual limitations, lack of explainability, real-time deployment constraints, and adversarial robustness. The review provides replicable evidence base and identifies actionable directions for the next generation of reliable, ethical, and deployable fake news detection systems.