A synthetic minority over-sampling–enhanced hybrid Indonesian bidirectional encoder representations from transformers–convolutional neural network model with cross-entropy optimization for imbalanced fake news detection in Indonesian online media

Abstract

The dissemination of fake news in Indonesian online media, particularly in political discourse, continues to increase and is often characterized by highly imbalanced class distributions. This study proposes a robust detection framework tailored for political fake news in the Indonesian language, with potential adaptability to other languages through appropriate linguistic and domain adjustments. The framework integrates a series of data mining and deep learning optimization strategies, emphasizing performance stability under severe data imbalance conditions. Specifically, we introduce a hybrid Indonesian Bidirectional Encoder Representations from Transformers-Convolutional Neural Network (IndoBERT-CNN) model enhanced through cross-entropy optimization, Word2Vec-based embedding refinement, and systematic imbalanced data handling using several Synthetic Minority Over-sampling (SMOTE) variants. Model performance was rigorously assessed using 10-fold cross-validation and benchmarked against multiple baseline architectures, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT), and Convolutional Neural Network—Long Short-Term Memory (CNN-LSTM). The experimental results show that the proposed model consistently outperforms all baseline models across various imbalanced scenarios, achieving a best F1-score of 90%, indicating its strong generalization capability and suitability for political fake news detection in Indonesian online media.