A Systematic Review of Machine Learning Approaches for Detecting Deceptive Activities on Social Media: Methods, Challenges, and Biases

The growing prevalence of misinformation, spam, and fake accounts on social media platforms such as Twitter, Facebook, and Instagram presents a serious threat to public discourse and safety. To address this, machine learning (ML) and deep learning (DL) models have been widely applied for detecting deceptive content. This systematic review analyzes 36 peer-reviewed studies published between January 2010 and July 2024 that apply ML/DL techniques to combat deception across major platforms. Among these, random forest and support vector machines were the most frequently used ML models (17 and 16 studies, respectively), while deep learning models such as artificial neural networks (8 studies) and LSTM (2 studies) also showed promising results. Performance metrics, including F1 scores and AUROC, were extracted when available, with several studies reporting F1 scores exceeding 90% for specific tasks like fake account detection. Despite these advances, 86% of studies inadequately addressed class imbalance, and only 39% consistently tuned hyperparameters. Furthermore, over 70% did not mention strategies for handling linguistic challenges such as negations. This review identifies key methodological limitations—including sample representativeness, inconsistent preprocessing, and reliance on accuracy in imbalanced settings—and offers recommendations to enhance generalizability, bias mitigation, and real-world applicability of ML-based misinformation detection.