<p>With rising living standards and growing public concern, food safety has become a critical issue, prompting regulatory systems to shift from manual, reactive approaches to intelligent, data-driven risk prediction. However, the multi-source, heterogeneous nature of food safety data poses significant challenges for traditional machine learning methods, which struggle to capture complex relational semantics and provide interpretable results. To address these limitations, we propose a visual analytics framework that integrates heterogeneous graph neural networks (HGNNs) with metapath-driven semantic modeling. By constructing a heterogeneous graph of food samples and attributes, our approach captures cross-type associations and models multi-relational structures with enhanced interpretability. A tailored visualization strategy further bridges data structures, embedding representations, and predictive behaviors, enabling users to intuitively explore attribute relationships, trace metapath influences, and understand risk prediction logic through interactive, multi-level visual analysis. Experiments on real-world food safety datasets demonstrate that our method significantly improves both prediction accuracy and model transparency, offering practical decision-making support for intelligent supervision and contamination prevention.</p> Graphical abstract <p></p>