On the Consistency of GNN Explainability Methods

Ehsan Hajiramezanali, Sepideh Maleki, Alex Tseng, Aicha BenTaieb, Gabriele Scalia, Tommaso Biancalani

Abstract: Despite the widespread utilization of post-hoc explanation methods for graph neural networks (GNNs) in high-stakes settings, there has been a lack of comprehensive evaluation regarding their quality and reliability. This evaluation is challenging primarily due to the data's non-Euclidean nature, arbitrary size, and complex topological structure. In this context, we argue that the consistency of GNN explanations, denoting the ability to produce similar explanations for input graphs with minor structural changes that do not alter their output predictions, is a key requirement for effective post-hoc GNN explanations. To fulfill this gap, we introduce a novel metric based on Fused Gromov--Wasserstein distance to quantify consistency. Finally, we demonstrate that current methods do not perform well according to this metric, underscoring the need for further research on reliable GNN explainability methods.

Submission Track: Full Paper Track

Application Domain: None of the above / Not applicable

Clarify Domain: Graph Neural Network

Survey Question 1: Graph Neural Networks are fundamental in drug discovery, and explainability would be very helpful in explaining the GNN predictions to experimentalists and building trust. To address this, several post-hoc GNN explainability methods have been proposed. However, there has been a lack of comprehensive evaluation regarding their quality and reliability, which is challenging due to the non-Euclidean nature, arbitrary size, and complex topological structure of the data. In this paper, we introduce a novel metric to quantify consistency.

Survey Question 2: Consistency of GNN explanations, denoting the ability to produce similar explanations for input graphs with minor structural changes that do not alter their output predictions, is a key requirement for effective post-hoc GNN explanations in drug discovery. In this paper, we introduce a novel metric based on Fused Gromov--Wasserstein distance to quantify this consistency. Our findings highlight that current methods do not perform well according to this metric, underscoring the need for further research on reliable GNN explainability methods.

Survey Question 3: We studied state-of-the-art GNN explainers, including SubgraphX, GstarX, PGExplainer, GNNexpainer, and GraphSVX. Some of these methods, such as SubgraphX and GraphSVX, are SHAP-based methods.

Submission Number: 57