SCLD-DHGAT: Dynamic Gumbel-Softmax Annealing and Semantic Center Loss for Multi-Class Fake News Detection

The rapid growth of misinformation on social media platforms has increased the demand for robust automated fake news detection systems. Although Graph Neural Networks (GNNs) have demonstrated strong capabilities in modeling complex relationships among news entities, most existing approaches focus on binary classification and employ static neighborhood structures, which can lead to over-smoothing, information loss, and unstable graph representation learning. In this paper, we propose an enhanced model based on a Decision-based Heterogeneous Graph Attention Network (DHGAT) for semi-supervised fake news detection. In the proposed model, a Dynamic Gumbel-Softmax Annealing mechanism is employed to gradually transition from exploration to exploitation during training, enabling more stable and reliable neighborhood selection. Then, a semantic center loss is incorporated to improve embeddings of the feature-space by reducing intra-class variation and increasing inter-class separability. Experiments conducted on the six-class LIAR benchmark dataset under semi-supervised settings with ۱۰%, ۲۰%, and ۳۰% labeled data demonstrate that the proposed model improves both classification performance and training stability. Using only ۳۰% labeled data, the model achieves a Macro F۱-score of ۰.۴۸۵۱, outperforming the baseline DHGAT model and several recent graph-based fake news detection methods. The statistical tests further indicate significant improvements over the evaluated comparison methods. The results demonstrate that combining adaptive temperature control with semantic representation regularization can improve the stability and classification performance of semi-supervised multi-class fake news detection.

کلیدواژه ها:

Social Networks ، Fake news detection ، graph neural networks ، Heterogeneous Graph Attention Networks ، semi-supervised learning