A Comprehensive Study of Bengali Fake News Detection Using Machine Learning, Deep Learning, and Transformer-Based Ensemble Approaches
Abstract
The rapid spread of misinformation on social media and online news portals makes
automated fake news detection a critical research problem, especially for low-resource
languages like Bengali. This thesis presents a leakage-aware pipeline that compares ML,
DL, and Transformer-based Ensemble models on the same dataset to identify the most
accurate and reliable approach. Al though many fake news detection studies already exist,
most focus on English or use only one type of model. In contrast, this research focuses on
Bengali and compares Machine Learning, Deep Learning, and Transformer-based
Ensemble models on the same dataset. The main objective is to identify the most accurate
and reliable approach for detecting Bengali fake news by collecting and preprocessing
Bengali news, training different AI models, evaluating their performance, and selecting
the best-performing model. A total of 27 models, including classical ML, DL,
BanglaBERT, XLM-RoBERTa, and eight ensemble strategies, were trained using TF-IDF
and stylometric features. SHAP, LIME, and attention analysis confirmed that the models
relied on genuine content rather than artifacts. By avoiding data leakage, the results
provide a more trustworthy performance estimate, while dataset size and label noise,
rather than model architecture, remain the main limitations. Among all 27 evaluated
models, the Soft-Voting Ensemble achieved the best overall performance with 87.09%
accuracy, 88.79% F1-score, and 93.96% ROC-AUC. It outperformed all individual
models by combining the strengths of diverse Machine Learning, Deep Learning, and
Transformer-based models. In complex situations, the model predicts based on learned
patterns. Since no AI model is 100% accurate, it should be used as a decision-support
tool. Transformer models improve context understanding, and Ensemble methods
combine multiple models to improve accuracy, making Bengali fake news detection more
effective
Collections
- 2026-2030 [51]