Machine Learning and Deep Learning for Detecting Fake News in a Low-Resource Language

Fake news detection has become a major problem in the digital age. This study presents an improved machine learning technique that achieves 91.99% accuracy in predicting fake news detection within Albanian textual datasets, demonstrating an improvement over existing baseline methodology. The implemented learning model uses 54 features that are specific to the Albanian language, such as red flags, credibility signals, punctuation patterns, and linguistic features. The model is tested on a balanced dataset of 3,994 news articles aggregated in Albanian from various sources. We compare it to several baselines, such as LSTM networks (80.35% accuracy) and BERT-augmented Naive Bayes classifiers (88.36% accuracy). In our Albanian dataset and experimental setting, XGBoost achieved 91.99% accuracy, indicating strong performance under the evaluated scenario.