A Comprehensive Review Of Machine Learning And Deep Learning Approaches For Detecting Fake News On Social Media

Abstract

In response to the escalating threat of fake news on social media, this systematic literature review analyzes the
recent advancements in machine learning and deep learning approaches for automated detection. Following
the PRISMA guidelines, we examined 90 peer-reviewed studies published between 2020 and 2024 to evaluate
the model effectiveness, identify limitations, and highlight emerging trends. Our analysis shows that deep
learning models, particularly transformer-based architectures such as BERT, consistently outperform
traditional machine learning methods, often achieving a high accuracy (Acc), precision (P), recall (R), and F1-
score (F1). For instance, a BERT-based model reported up to 89.9% accuracy on the Kaggle fake news
dataset and above 98% accuracy on other public datasets, including ISOT, Fake-orReal, and D3. Similarly,
the GANM model demonstrated robust performance on the FakeNewsNet dataset by integrating text and
social features. Transfer learning and multimodal models that incorporate user behaviour and network
information significantly improve detection in diverse, low-resource environments. However, challenges
persist in terms of the dataset quality, model interpretability, domain generalisability, and realtime
deployment. This review also underscores the limited adoption of few-shot and zero-shot learning techniques,
highlighting a promising direction for future research on handling emerging misinformation using minimal
training data. To support practical deployment, we advocate the development of explainable, multilingual, and
lightweight models with greater emphasis on human-centred evaluation and ethical considerations. Our
findings provide a foundation for researchers and practitioners to build scalable, trustworthy, and contextaware
fake news detection systems for global use.