Graph Neural Network Causal Explanation via Neural Causal Models (2407.09378v1)

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

• The paper introduces CXGNN, a causal explainer that integrates causal inference principles to differentiate causal and non-causal graph substructures.
• It employs Neural Causal Models optimized through gradient descent to compute cause-effect relationships efficiently within complex graphs.
• Empirical studies on synthetic and real-world datasets show that CXGNN significantly outperforms traditional methods in unveiling accurate groundtruth causal explanations.

Graph Neural Network Causal Explanation via Neural Causal Models

The paper "Graph Neural Network Causal Explanation via Neural Causal Models" presents an innovative approach to the challenge of explaining the predictions made by Graph Neural Networks (GNNs). Traditionally, explaining GNNs has relied predominantly on associating predictions with subgraphs believed to hold maximum predictive power. These approaches, however, are susceptible to spurious correlations. The authors introduce a GNN causal explainer, termed CXGNN, which leverages causal inference principles to offer explanations that more accurately reflect the underlying causal mechanisms of GNN predictions.

Key Contributions

1. Causal Structure and Structural Causal Model (SCM): The authors propose that each graph consists of causal and non-causal subgraphs. They define a causal structure for a given graph, on which they build the corresponding Structural Causal Model (SCM). This structure allows the computation of cause-effect relationships among nodes via interventions.
2. Neural Causal Model (NCM): Recognizing the computational complexity inherent in calculating cause-effect in real-world graphs, the paper introduces Neural Causal Models (NCMs). These models are a specialized, trainable form of SCM, inspired by recent advancements in neural causal models, allowing for gradient descent-based optimization. The authors provide proofs of concept for constructing these GNN-specific NCMs.
3. Causal Explanation via CXGNN: By training the GNN-NCMs, the explainer uncovers the subgraph that causally explains the GNN predictions. Evaluation on synthetic and real-world datasets revealed that CXGNN significantly outperforms existing GNN explainers in discovering exact groundtruth explanations.

Evaluation and Findings

The paper’s evaluation on multiple synthetic and real-world datasets provides strong support for its claims. CXGNN demonstrated significant superiority over traditional association-based state-of-the-art explainers, including GNNExplainer, PGMExplainer, and others in terms of correctly identifying groundtruth explanatory subgraphs. Particularly, the groundtruth match accuracy observed in the results attests to the robustness of CXGNN in uncovering causal explanations, a feat where previous explainers often falter.

Implications and Future Directions

The implications of this paper are twofold. Practically, CXGNN offers a more reliable tool for understanding and interpreting GNN predictions. Theoretically, it challenges the current paradigm in GNN explanation methods by rooting its approach in causal inference, thus opening up new avenues for research in both explainability and causal discovery in machine learning.

Going forward, potential developments might explore the scalability of CXGNN to larger graph structures and its applicability across different domains of graph data, such as social networks and biological networks. Additionally, integrating CXGNN with adversarial robustness techniques could fortify its efficacy in real-world applications.

In sum, the paper provides a detailed methodology and solid empirical results that make a compelling case for adopting causal models in the explanation of GNN predictions, aligning the explanation approach more closely with true causal mechanisms. This work sets a foundation for future innovations in the interpretability of complex neural network architectures.