A Multi Stage Ensemble Framework for Fake News Detection: Integrating Traditional Machine Learning, Deep Learning, and Advanced Feature Engineering
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Objective: The rapid spread of online misinformation demands scalable and accurate detection systems. This study proposes a hierarchical multi-stage ensemble framework that integrates traditional machine learning, deep learning, and multi-level feature engineering for binary fake news classification. Methods: A balanced corpus of 44,898 news articles (fake vs. real) was constructed with body text and ground-truth labels. Preprocessing employed a three-stage text cleaning pipeline followed by four-level feature engineering capturing lexical, syntactic, semantic, and stylometric characteristics, augmented with TF IDF representations and mutual-information-based feature selection. The modelling pipeline comprises four stages: (i) traditional classifiers (Random Forest, SVM, Naive Bayes); (ii) advanced classifiers (Gradient Boosting, Logistic Regression, Multi-Layer Perceptron) trained on original features plus Stage 1 probability estimates; (iii) deep learning models (LSTM, CNN, and hybrid LSTM+handcrafted features); and (iv) a calibrated Random Forest meta-learner operating on probabilistic outputs of all nine upstream models. Performance was evaluated using stratified train-test splits, cross-validation, accuracy, precision, recall, F1-score, ROC-AUC, ablation studies, and McNemar's tests. Results: Traditional classifiers with TF IDF and handcrafted features achieved up to 99.78% accuracy (Random Forest). Stage 2 models further improved performance, with Gradient Boosting reaching 99.71% accuracy. Deep learning models achieved 95.23–96.47% accuracy, confirming the complementary value of sequence-based representations. The calibrated Stage 4 meta-ensemble achieved 99.83% accuracy, 99.83% F1-score, and 0.9997 ROC-AUC, significantly outperforming all individual models. Ablation experiments confirmed each stage and feature level contributed positively. Conclusion: The proposed four-stage hierarchical ensemble provides a highly accurate, modular framework for fake news detection, offering a strong baseline for future research including transformer-based encoders and cross-domain validation.
Fields of Study & Subjects
Misinformation and Its Impacts
Spam and Phishing Detection
Text and Document Classification Technologies
Publication details
- Journal / source
- International Journal of Scientific Research and Management (IJSRM)
- Volume
- 14
- Issue
- 03
- Pages
- 2802–2817
- Publication date
- 1 January 2026
- Language
- EN