GroupEx: Toward Group-Level Explanations of Graph Neural Networks
提出群体级解释范式GroupEx，通过聚类图嵌入并优化同构目标，发现共享同一解释的图子群，比现有方法提供更深入的结构化理解。
The increasing adoption of graph neural networks (GNNs) in critical real-world applications brings with it theneed for interpretability and explainability of these models. The existing paradigms of GNN explainability typically operate at the extremes. Instance-level methods offer a microscopic view focused on individual predictions, whereas model-level approaches provide a macroscopic summary of the classifier behavior on a target class. We introduce the group-level paradigm, which views explanations at the right resolution, grounded in the intuition that groups of instances embedded close together in the latent space share a common rationale for belonging to the same target class. We show that the instance-level and model-level paradigms are the special cases of this broader framework. To realize this paradigm, we propose GroupEx, a novel method for interpreting GNNs at the group level. GroupEx has the capability to identify the subgroups of graphs within a target class using a group identity (GI) score. It employs a group extractor that can extract a subgroup of graphs that share a common explanation for their class identity by clustering graph embeddings in the latent space of the classifier. A specially devised explanation generator can then discover a group-level explanation by optimizing a novel isomorphism objective, which ensures that the discovered explanation commonly explains the class identity of all the instances in the group. Empirical results on real-world and synthetic datasets demonstrate that GroupEx provides deeper insights into GNN decision-making than the existing state-of-the-art explainability methods, enabling a more structured and interpretable understanding of model predictions.