Enhancing Arabic Fake News Detection: An Empirical Study on the GAFND Benchmark and the Impact of Feature Engineering

The advancement of Arabic fake news detection is hindered by a lack of large-scale, unified datasets. To address this, we introduce the Global Arabic Fake News Dataset (GAFND), a new corpus of 382,701 entries. This study investigates whether models trained on textual content can be improved by incorporating engineered features. Using a rigorous 5-fold stratified cross-validation methodology, we benchmarked a wide array of machine learning classifiers, comparing performance on "text-only" data versus a combination of text and 19 engineered features. Our results show that the inclusion of these features provides a significant and consistent performance boost across nearly all models. Gradient boosting classifiers like CatBoost and LightGBM demonstrated the most substantial gains, with accuracy improving by up to 11 percentage points. Feature importance analysis confirmed that metrics capturing article length, complexity, and style are highly predictive. This work contributes a valuable public dataset and, through this robust empirical validation, proves that a hybrid feature approach significantly enhances Arabic fake news detection, establishing a new and statistically sound set of performance benchmarks.