Enhancing Arabic Fake News Detection: Evaluating Data Balancing Techniques Across Multiple Machine Learning Models

The spread of fake news has become a serious concern in the era of rapid information dissemination through social networks, especially when it comes to Arabic-language content, where automated detection systems are not as advanced as those for English-language content. This study evaluates the effectiveness of various data balancing techniques, such as class weights, random under-sampling, SMOTE, and SMOTEENN, across multiple machine learning models, namely XGBoost, Random Forest, CNN, BIGRU, BILSTM, CNN-LSTM, and CNN-BIGRU, to address the critical challenge of dataset imbalance in Arabic fake news detection. Accuracy, AUC, precision, recall, and F1-score were used to evaluate the performance of these models on balanced and imbalanced datasets. The results show that SMOTEENN greatly improves model performance, especially the F1-score, precision, and recall. In addition to advancing the larger objective of preserving information credibility on social networks, this study emphasizes the need for advanced data balancing strategies to improve Arabic fake news detection systems.