A Novel Approach to GNN Explainability: Distilling Knowledge With Inter-Layer Alignment.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Science

Background:

• Graph Neural Networks (GNNs) excel at network data analysis but suffer from a "black-box" problem, hindering trust and application.
• Existing GNN explanation methods are often complex and costly due to reliance on subgraph selection and combinatorial optimization.
• Over-smoothing in GNNs further complicates model interpretability and explanation generation.

Purpose of the Study:

• To develop a lower-complexity proxy model for explaining GNN decision-making processes.
• To enhance the transparency and trustworthiness of GNN models in complex network analysis.
• To address the challenges of high explanation costs and the impact of over-smoothing on GNN interpretability.

Main Methods:

• Introduced a proxy model derived from complex GNNs using knowledge distillation.
• Employed inter-layer alignment during distillation to ensure proxy model fidelity to the original GNN.
• Theoretically proved the faithfulness of explanations generated by the proxy model to both models.

Main Results:

• The proposed method effectively distills insights from complex GNNs into a manageable proxy model.
• Inter-layer alignment successfully mitigates over-smoothing effects, improving explanation quality.
• Experimental results on real-world datasets demonstrate the effectiveness and robustness of the proposed explanation technique.

Conclusions:

• The developed proxy model offers a more transparent and efficient approach to explaining GNNs.
• Knowledge distillation with inter-layer alignment is a viable strategy for enhancing GNN interpretability.
• The method provides faithful and robust explanations, paving the way for broader GNN adoption.