Balancing the truth: social media misinformation detection via GraphSAGE and feature diversity

Abstract

Misinformation on social media platforms poses significant risks to public opinion and decision-making due to its rapid dissemination. Traditional text-based approaches often struggle to capture the complex interactions between textual content, user behavior, and contextual relationships. To address this limitation, we propose a heterogeneous graph-based misinformation detection framework integrating user, tweet, and hashtag relations with diverse text representations. In our design, tweets are encoded using both sparse lexical features (Term Frequency–Inverse Document Frequency (TF-IDF)) and dense semantic embeddings (Sentence-Bidirectional Encoder Representations from Transformers (SBERT)), while classification is performed through multiple Graph Neural Network (GNN) architectures, namely Graph Sample and Aggregate (GraphSAGE), Graph Convolutional Network (GCN), and Graph Attention Network (GAT). To account for potential class imbalance, class-weighted loss functions are incorporated during training. Experiments on the TruthSeeker dataset demonstrate that GraphSAGE consistently outperforms GCN and GAT across all feature settings. In particular, the best-performing configuration—TF-IDF combined with GraphSAGE—achieves 94% Accuracy and 94% F1-score, with comparable performance under both standard and class-weighted training. Analysis of training dynamics indicates stable convergence across configurations; GraphSAGE-based models continue to benefit from longer training, reaching 96.73% accuracy at 200 epochs. Comparative evaluations suggest that modeling heterogeneous relations with diverse feature representations provides advantages over purely text-based baselines within the examined setting. These results highlight the value of integrating relational graph structures with complementary feature representations for misinformation detection. While class-weighted training does not substantially alter performance under the near-balanced dataset distribution, it maintains stable learning across classes without degrading overall results. The proposed framework provides a foundation for future extensions, including federated learning, explainability techniques, and cross-domain applications in multilingual contexts.