Reinforced Causal Explainer for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Explainability is critical for understanding graph neural network (GNN) predictions.
• Feature attribution methods highlight explanatory subgraphs but often assume edge independence, leading to unfaithful and verbose explanations.
• Existing methods overlook the coalition effect and dependencies among edges.

Purpose of the Study:

• To address the limitations of existing feature attribution methods in GNN explainability.
• To develop a method that accounts for edge dependencies and coalition effects for more faithful explanations.
• To generate concise and generalizable explanations for GNNs.

Main Methods:

• Proposes Reinforced Causal Explainer (RC-Explainer), a reinforcement learning agent.
• Frames explanation as a sequential decision process of constructing explanatory subgraphs by adding edges.
• Employs a policy network for edge addition and a reward function quantifying causal effect, considering edge dependencies and coalition.

Main Results:

• RC-Explainer generates faithful and concise explanations by accounting for edge dependencies.
• The method demonstrates better generalization power to unseen graphs.
• Achieves comparable or superior performance to state-of-the-art methods on graph classification datasets regarding predictive accuracy and contrastivity.

Conclusions:

• RC-Explainer effectively overcomes the limitations of assuming edge independence in GNN explainability.
• The reinforcement learning approach provides a robust framework for generating high-quality explanations.
• The proposed method enhances GNN interpretability and trustworthiness.