Explainable Fake News Detection in Vietnamese Media Using Content Semantics, Source Reliability, and Information Diffusion Patterns

With the development of social networks and digital media, the spread of misinformation and fake news in Vietnam is becoming increasingly serious. Existing fake news detection systems mainly focus on content analysis without considering other important aspects such as source credibility and information dissemination behavior. In particular, most previously published studies have focused on English data, with few studies on less widely used languages like Vietnamese. Furthermore, these studies lack explanatory mechanisms that allow users to understand the model's decisions. This study proposes a combined approach to explain how Vietnamese fake news is detected by integrating semantic content analysis, source credibility assessment, and information dissemination models. The proposed solution combines PhoBERT-based semantic embedding, source credibility scoring, and graph neural networks to model dissemination through a multimodal, attention-based combined architecture. Furthermore, explainable artificial intelligence techniques, including SHAP analysis and attention visualization, were integrated to improve explainability. A dataset of Vietnamese fake news was constructed from online newspapers and social media platforms, including dissemination metadata and source information. Experimental results showed that the proposed model significantly outperformed traditional machine learning, deep learning, and transformer-based baseline models in terms of Accuracy, Precision, Recallability, F1 score, and ROC-AUC. This study contributes a framework for disinformation detection that is explainable and recognizes dissemination, specifically designed for the Vietnamese media ecosystem.