Using the WELFake dataset—a benchmark containing 44, 898 labeled news articles—the authors evaluated the performance of Logistic Regression (LR), Decision Tree (DT), and Naive Bayes (MNB) classifiers. The study employed TF-IDF vectorization for feature extraction and utilized 10-fold cross-validation to ensure unbiased results Key Findings: Superior Performance: Logistic Regression emerged as the most effective model, achieving a mean accuracy of 96. 51%. Comparative Analysis: LR outperformed the Naive Bayes baseline (90. 21%) and significantly surpassed the Decision Tree algorithm (88. 33%). Statistical Significance: An Independent Sample T-Test confirmed that LR’s 8. 18-point accuracy advantage over DT is statistically significant (p < 0. 001). Reliability: LR demonstrated a much lower False Positive Rate (3. 04% vs. 10. 67% for DT), making it a safer choice for real-world social media content moderation where misclassifying legitimate news is a high risk. Conclusion: The research concludes that Logistic Regression is more robust for high-dimensional, sparse text data than Decision Trees, which are prone to overfitting in such environments. The authors suggest that while LR is ideal for production deployment due to its efficiency and well-calibrated probability estimates, future work should explore multi-modal features and contextual language models like BERT.
Raj et al. (Tue,) studied this question.