Graph Neural Network Causal Explanation via Neural Causal Models

TL;DR

This work tackles the problem of explainable graph neural networks by shifting from association-based explanations to causal explanations. It introduces CXGNN, a framework that builds a graph-level causal structure (SCM) around a reference node and learns a trainable GNN-NCM constrained by this structure to quantify cause–effect relations and identify a causal explanatory subgraph that drives predictions. Empirical results on synthetic and real-world datasets show CXGNN significantly improves exact groundtruth explanation identification compared with state-of-the-art baselines, though exact matches remain challenging in real-world data due to groundtruth ambiguity. Overall, CXGNN provides a principled, scalable approach to causal explanation in graphs with potential impact on robust interpretation and decision-making in graph-centric tasks.

Abstract

Graph neural network (GNN) explainers identify the important subgraph that ensures the prediction for a given graph. Until now, almost all GNN explainers are based on association, which is prone to spurious correlations. We propose {\name}, a GNN causal explainer via causal inference. Our explainer is based on the observation that a graph often consists of a causal underlying subgraph. {\name} includes three main steps: 1) It builds causal structure and the corresponding structural causal model (SCM) for a graph, which enables the cause-effect calculation among nodes. 2) Directly calculating the cause-effect in real-world graphs is computationally challenging. It is then enlightened by the recent neural causal model (NCM), a special type of SCM that is trainable, and design customized NCMs for GNNs. By training these GNN NCMs, the cause-effect can be easily calculated. 3) It uncovers the subgraph that causally explains the GNN predictions via the optimized GNN-NCMs. Evaluation results on multiple synthetic and real-world graphs validate that {\name} significantly outperforms existing GNN explainers in exact groundtruth explanation identification

Figures (18)

• Figure 1: Visualizing explanation results (subgraph containing the red nodes) by our CXGNN on synthetic graphs.
• Figure 2: Loss curves of training the GNN-NCMs on the groundtruth nodes ( green curves) and non-groundtruth ones ( red curves) on two random chosen graphs from BA+House. More examples in other datasets are shown in Appendix \ref{['app:exp']}.
• Figure 3: Node expressivity distributions on two unsuccessful graphs from BA+Cycle. Green bars correspond to nodes that are in the groundtruth, while red bars correspond to nodes that are not. More examples in other datasets are shown in Appendix \ref{['app:exp']}.
• Figure 4: Explanation results (subgraph containing the yellow nodes) by our CXGNN on real-world graphs. The left and right two graphs are in Benzene and F.C., respectively.
• Figure 5: Loss curves of training the GNN-NCMs on the groundtruth nodes ( green curves) and non-groundtruth ones ( red curves) on two graphs from the two real-world datasets, respectively. More examples are shown in Appendix \ref{['app:exp']}.

Theorems & Definitions (20)

• definition thmcounterdefinition: Intervention and Causal Effects
• definition thmcounterdefinition: Causal structure of a graph
• theorem thmcountertheorem: GNN-SCM
• definition thmcounterdefinition: $\mathcal{G}$-Constrained GNN-NCM (constructive)
• theorem thmcountertheorem: GNN-NCM
• theorem thmcountertheorem: Node explainability
• theorem thmcountertheorem: Explainable node expressivity
• definition thmcounterdefinition: PCH layers
• definition thmcounterdefinition: $\mathcal{G}$-Consistency
• definition thmcounterdefinition: $\mathcal{G}$-Constrained NCM