Explainable Machine Learning-Based Fake News Detection Using NLP Techniques

Abstract

The rapid spread of misinformation on social media poses significant societal risks. This study develops an interpretable and accurate fake-news detection system that combines traditional supervised learning with Explainable AI (XAI). Two large, publicly available datasets were merged into a single corpus and preprocessed (cleaning, tokenization, stop-word removal, and lemmatization). Features were extracted with TF–IDF and n-gram vectorization (uni/bi-grams). We trained Support Vector Machine (SVM), Logistic Regression, and a Passive-Aggressive classifier, and evaluated them on a held-out test set. The SVM achieved the highest accuracy (93.5%), outperforming Logistic Regression (91.2%) and the Passive-Aggressive classifier (90.5%); precision, recall, and F1-score similarly supported this ranking. Confusion-matrix analysis indicated balanced detection across fake and real classes. To improve transparency, we applied SHAP and LIME to highlight influential tokens and n-grams for both global trends and individual predictions, enabling users to verify the cues driving classifications. These results show that pairing efficient linear text models with post-hoc explanations can deliver competitive accuracy while mitigating black-box concerns. Limitations include vulnerability to evolving adversarial writing styles, limited multilingual coverage, and the need for real-time inference. Future work will explore continual/online learning, cross-lingual transfer, and streaming deployment to enhance robustness and broader applicability.