Applications of machine learning for COVID-19 misinformation: a systematic review

A. R. Sanaullah, Anupam Das, Anik Das, Muhammad Ashad Kabir, Kai Shu

Abstract: The inflammable growth of misinformation on social media and other platforms during pandemic situations like COVID-19 can cause significant damage to the physical and mental stability of the people. To detect such misinformation, researchers have been applying various machine learning (ML) and deep learning (DL) techniques. The objective of this study is to systematically review, assess, and synthesize state-of-the-art research articles that have used different ML and DL techniques to detect COVID-19 misinformation. A structured literature search was conducted in the relevant bibliographic databases to ensure that the survey was solely centered on reproducible and high-quality research. We reviewed 43 papers that fulfilled our inclusion criteria out of 260 articles found from our keyword search. We have surveyed a complete pipeline of COVID-19 misinformation detection. In particular, we have identified various COVID-19 misinformation datasets and reviewed different data processing, feature extraction, and classification techniques to detect COVID-19 misinformation. In the end, the challenges and limitations in detecting COVID-19 misinformation using ML techniques and the future research directions are discussed.

Defining the Infodemic

The COVID-19 pandemic triggered a secondary crisis: an "infodemic." The rapid spread of rumors, conspiracy theories, and fabricated news on social media didn't just cause confusion; it led to real-world harm, such as medical overdoses and vaccine hesitancy. While researchers rushed to develop automated detection systems, the literature became fragmented.

A critical first step in addressing this challenge is establishing a clear taxonomy. As shown in the figure below, "misinformation" serves as an umbrella term encompassing various distinct categories, each requiring different detection strategies.

Taxonomy of misinformation types identified in the review.

Methodology

To synthesize the current state of the field, a structured literature search was conducted across major databases including Scopus, Web of Science, and Google Scholar. The review followed a rigorous selection process to ensure only high-quality, reproducible research was included.

Systematic review process for article selection.

Out of 260 initial articles, 43 papers met the final inclusion criteria. These studies were analyzed to map the complete detection pipeline, from preprocessing and feature extraction to final classification architectures.

The Detection Pipeline

The authors reviewed the complete technical pipeline used in state-of-the-art detection:

  • Preprocessing: Standard NLP techniques like tokenization and stop-word removal are used alongside specialized steps like POS (Part-of-Speech) tagging and lemmatization.
  • Feature Extraction: Approaches range from traditional frequency-based methods (TF-IDF, Bag-of-Words) to psycholinguistic features (LIWC) and modern word embeddings (Word2Vec, GloVe).
  • Classification Architectures:
    • Traditional ML: Support Vector Machines (SVM), Logistic Regression (LR), and Random Forests (RF).
  • Deep Learning: Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRU).
  • Transformers: The current "gold standard," including BERT, RoBERTa, and domain-specific variants like COVID-Twitter-BERT (CT-BERT).

Dataset Landscape

The review identifies several key datasets, primarily sourced from Twitter, Weibo, and fact-checking sites. Notable repositories include CoAID and ReCOVery for general news, MM-COVID for multilingual data, and ArCOV-19 for Arabic-language rumors.

However, a significant challenge remains: many datasets are heavily imbalanced, containing far more "real" news than "fake" news, which can skew model performance. Furthermore, there is a clear trend moving away from manual feature engineering toward automated feature extraction via neural embeddings and transformer layers.

Key Insights and Performance

  • Domain Adaptation: Models pre-trained on COVID-specific corpora (like CT-BERT) consistently outperform generic models. Contextual nuances in pandemic-related language are critical for accuracy.
  • Deep Learning Dominance: DL models are significantly more efficient at capturing complex, non-linear patterns. For instance, a CNN-based model achieved 99.99% accuracy on the COVID-19-FAKES dataset, while BERT-based models frequently hit the 99% mark.
  • Traditional Baselines: Surprisingly, simpler models like SVM and Logistic Regression achieved 93-96% accuracy in several studies, suggesting they remain viable for binary classification tasks where features are well-chosen.

Limitations and Future Directions

Despite high accuracy scores, critical gaps remain:

  • Multimodality: Most research focuses on text, ignoring the rise of visual misinformation (memes, edited videos).
  • Multilingual Gap: The vast majority of studies focus on English, leaving non-English speaking populations vulnerable.
  • Explainability: Most DL models are "black boxes." Future work must focus on Explainable AI (XAI) so moderators understand why content was flagged.
  • Unsupervised Learning: As labeling data is slow, moving toward semi-supervised or reinforcement learning is essential to keep pace with the rapid growth of misinformation.

Summary updated

Abstract

The inflammable growth of misinformation on social media and other platforms during pandemic situations like COVID-19 can cause significant damage to the physical and mental stability of the people. To detect such misinformation, researchers have been applying various machine learning (ML) and deep learning (DL) techniques. The objective of this study is to systematically review, assess, and synthesize state-of-the-art research articles that have used different ML and DL techniques to detect COVID-19 misinformation. A structured literature search was conducted in the relevant bibliographic databases to ensure that the survey was solely centered on reproducible and high-quality research. We reviewed 43 papers that fulfilled our inclusion criteria out of 260 articles found from our keyword search. We have surveyed a complete pipeline of COVID-19 misinformation detection. In particular, we have identified various COVID-19 misinformation datasets and reviewed different data processing, feature extraction, and classification techniques to detect COVID-19 misinformation. In the end, the challenges and limitations in detecting COVID-19 misinformation using ML techniques and the future research directions are discussed.
BibTeX
@inproceedings{sanaullah2022applications,
  author = {A. R. Sanaullah and Anupam Das and Anik Das and Muhammad Ashad Kabir and Kai Shu},
  title = {{Applications of machine learning for COVID-19 misinformation: a systematic review}},
  booktitle = {Soc. Netw. Anal. Min},
  year = {2022},
  url = {https://openreview.net/forum?id=302I3Y9f45},
}
Figure