Abstract
<title>Abstract</title> <p>The rapid expansion of digital media and online social networking platforms has significantly increased the dissemination of fake news, creating serious social, political, and economic challenges worldwide. Detecting fake news accurately remains difficult because deceptive information often resembles authentic news in both linguistic style and contextual presentation. This study presents a secondary-data meta-analysis that examines the effectiveness of feature representation techniques employed in machine learning-based fake news detection. Drawing upon benchmark datasets and previous empirical studies reviewed in the source thesis, the analysis compares traditional representations, including Term Frequency–Inverse Document Frequency (TF-IDF), N-gram, Bag-of-Words, Word2Vec, and transformer-based contextual embeddings, together with various machine learning and ensemble classifiers. The comparative findings indicate that hybrid feature representations generally outperform individual feature extraction methods by capturing both term importance and contextual relationships among words. Furthermore, ensemble learning approaches consistently demonstrate greater predictive stability than individual classifiers across different datasets. The study concludes that combining efficient feature engineering with lightweight machine learning models provides a practical and computationally efficient solution for fake news detection, while highlighting research directions toward more robust, multilingual, and explainable detection frameworks.</p>