Title: HiSE: A Lightweight Hierarchical Semantic Explainer for Heterogeneous Graph Neural Networks

Abstract:

While Heterogeneous Graph Neural Networks (HGNNs) have shown exceptional capability in handling complex relational data, their lack of interpretability poses a significant barrier to adoption in high-stakes scenarios. Current explanation techniques are hindered by two primary drawbacks: first, they often overlook the intrinsic semantic hierarchy of HGNNs, which compromises the fidelity of the explanations relative to the model’s internal logic; second, feature attribution methods frequently depend on computationally intensive search or perturbation strategies, resulting in poor efficiency. To overcome these challenges, we introduce HiSE, a lightweight, feature-focused interpretable framework designed for HGNNs. HiSE generates semantically informed feature explanations by leveraging hierarchical semantic modeling. Specifically, at the semantic level, it utilizes local surrogate models grounded in the Least Absolute Shrinkage and Selection Operator (LASSO) to derive sparse feature representations within each semantic view. At the cross-semantic level, it dynamically assesses the contribution of various semantic views using KL divergence to synthesize a cohesive explanation. Comprehensive experiments confirm that HiSE surpasses current state-of-the-art methods in fidelity, robustness, and cross-semantic explanatory power. Furthermore, its efficient, lightweight architecture ensures low computational costs, making it suitable for deploying on large-scale, intricate real-world heterogeneous graphs.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC