CiRLExplainer: Causality-Inspired Explainer for Graph Neural Networks via Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are powerful tools for data analysis but often lack transparent explanations for their predictions.
• Existing explainability methods struggle to account for complex causal relationships within graph structures.

Purpose of the Study:

• To introduce CiRLExplainer, a novel GNN explainability model based on causal attribution.
• To provide precise semantic explanations for GNN predictions by analyzing causal relationships.

Main Methods:

• Constructing a causal graph to identify relationships between graph structure and GNN predictions.
• Employing a backdoor adjustment strategy to handle confounding factors (node attributes).
• Utilizing reinforcement learning for sequential edge selection to build explanatory subgraphs.

Main Results:

• CiRLExplainer outperforms state-of-the-art explanation techniques in accuracy (ACC) and area under the curve (AUC) metrics.
• Experimental validation confirms the effectiveness of addressing node attributes as confounding factors.
• The model achieved significant AUC improvements (5.89%, 5.69%, 4.87%) over baseline models on various datasets.

Conclusions:

• CiRLExplainer provides a versatile and effective causal attribution framework for GNN explainability.
• The model enhances GNN interpretability and predictive accuracy through a principled causal approach.
• Future work can explore further applications of causal inference in GNN explainability.