A HYBRID BERT–GNN FOR DETECTING HOAXES AND NEGATIVE CONTENT IN INDONESIAN SOCIAL MEDIA

The rapid spread of hoaxes on social media threatens public trust and information integrity, especially within the Indonesian digital landscape. This study proposes a hybrid deep learning model that integrates transformer-based semantic representation from IndoBERT with Graph Neural Networks (GNNs) to enhance hoax detection performance. A heterogeneous social graph is constructed to model relationships among posts, users, and news sources, where post node features are extracted from the [CLS] embeddings of a fine-tuned IndoBERT. The GNN component consists of two graph convolutional layers with ReLU activation and dropout, followed by a multilayer perceptron classifier for binary classification. Experiments conducted on the Indonesia False News dataset (Kaggle) employ SMOTE resampling to handle class imbalance and 5-fold stratified cross-validation for robust evaluation across three configurations: BERT-only, GNN-only, and the proposed BERT–GNN hybrid model. The hybrid model achieves an average F1-score of 0.89 ± 0.01 and ROC-AUC of 0.92 ± 0.01, outperforming both single-model baselines while maintaining a balanced precision–recall trade-off. These results confirm that combining contextual semantic understanding with relational graph topology substantially enhances accuracy, robustness, and generalization in detecting hoaxes within Indonesian-language social media content