HiSE: A Lightweight Hierarchical Semantic Explainer for Heterogeneous Graph Neural Networks
Researchers have developed HiSE, a new interpretable model designed for heterogeneous graph neural networks (HGNNs). This lightweight approach addresses the challenge of explaining HGNN decisions in critical applications by reflecting the model's semantic hierarchy. HiSE uses LASSO for sparse feature representations within semantic views and KL divergence to unify explanations across these views, outperforming existing methods in fidelity and efficiency. AI
IMPACT Improves the explainability of complex graph neural networks, crucial for high-stakes applications.