An Explainable Ensemble Learning Framework for Accurate Fake News Detection Using TF-IDF Features

Abstract

The rapid proliferation of misinformation on digital platforms has become a critical societal challenge. This paper presents a comprehensive comparative study of four gradient boosting ensemble classifiers — Gradient Boosting Classifier (GBC), XGBoost, LightGBM, and CatBoost — for automated fake news detection. The publicly available Kaggle Fake and Real News Dataset  is used as the benchmark. A systematic text preprocessing pipeline is applied, followed by Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction using unigrams and bigrams. All four models are trained on an 80/20 stratified split and evaluated using Accuracy, Precision, Recall, F1-Score, ROC-AUC, and RMSE. LightGBM achieves the highest F1-Score of 0.9959 and ROC-AUC of 0.9996, followed closely by Gradient Boosting (F1=0.9958), XGBoost (F1=0.9956), and CatBoost (F1=0.9955). LIME (Local Interpretable Model-Agnostic Explanations) analysis is applied to the best model to identify the most influential textual features driving predictions. Error analysis through misclassification counts further validates the robustness of all classifiers. The results confirm that TF-IDF-based gradient boosting ensembles provide highly accurate, efficient, and interpretable solutions for real-world fake news detection.