Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Neighbor Explainability for Graph Neural Networks

Oscar Llorente, Rana Fawzy, Jared Keown, Michal Horemuz, Péter Vaderna, Sándor Laki, Roland Kotroczó, Rita Csoma, János Márk Szalai-Gindl

TL;DR

This paper studies neighbor-level explainability for node classification in Graph Neural Networks and introduces four metrics (Loyalty, Inverse Loyalty, Loyalty Probabilities, Inverse Loyalty Probabilities) to evaluate how well explanations identify important neighbors. It reformulates several explainability methods to output per-neighbor importances and compares gradient-based approaches (Saliency Map, Deconvnet, Guided-Backpropagation, SmoothGrad) with graph-specific explainers (GNNExplainer, PGExplainer) across GCN and GAT models on Cora, CiteSeer, and PubMed. The results show gradient-based explanations yield similar performance—unlike typical CV findings—while GNNExplainer is best for highly influential neighbors; PGExplainer generally underperforms, and self-loops critically affect performance, with gradient-based methods faltering when self-loops are absent. These findings guide practitioners in selecting explanations for neighbor-level interpretation in GNNs and suggest avenues for future work to address limitations when self-loops are missing and to understand why gradient-based methods tend to converge toward similar results.

Abstract

Explainability in Graph Neural Networks (GNNs) is a new field growing in the last few years. In this publication we address the problem of determining how important is each neighbor for the GNN when classifying a node and how to measure the performance for this specific task. To do this, various known explainability methods are reformulated to get the neighbor importance and four new metrics are presented. Our results show that there is almost no difference between the explanations provided by gradient-based techniques in the GNN domain. In addition, many explainability techniques failed to identify important neighbors when GNNs without self-loops are used.

Evaluating Neighbor Explainability for Graph Neural Networks

TL;DR

This paper studies neighbor-level explainability for node classification in Graph Neural Networks and introduces four metrics (Loyalty, Inverse Loyalty, Loyalty Probabilities, Inverse Loyalty Probabilities) to evaluate how well explanations identify important neighbors. It reformulates several explainability methods to output per-neighbor importances and compares gradient-based approaches (Saliency Map, Deconvnet, Guided-Backpropagation, SmoothGrad) with graph-specific explainers (GNNExplainer, PGExplainer) across GCN and GAT models on Cora, CiteSeer, and PubMed. The results show gradient-based explanations yield similar performance—unlike typical CV findings—while GNNExplainer is best for highly influential neighbors; PGExplainer generally underperforms, and self-loops critically affect performance, with gradient-based methods faltering when self-loops are absent. These findings guide practitioners in selecting explanations for neighbor-level interpretation in GNNs and suggest avenues for future work to address limitations when self-loops are missing and to understand why gradient-based methods tend to converge toward similar results.

Abstract

Explainability in Graph Neural Networks (GNNs) is a new field growing in the last few years. In this publication we address the problem of determining how important is each neighbor for the GNN when classifying a node and how to measure the performance for this specific task. To do this, various known explainability methods are reformulated to get the neighbor importance and four new metrics are presented. Our results show that there is almost no difference between the explanations provided by gradient-based techniques in the GNN domain. In addition, many explainability techniques failed to identify important neighbors when GNNs without self-loops are used.
Paper Structure (17 sections, 6 equations, 4 figures, 5 tables)

This paper contains 17 sections, 6 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Explainability methods
  3. Saliency Map
  4. Deconvnet and Guided-Backpropagation
  5. SmoothGrad
  6. GNNExplainer
  7. PGExplainer
  8. Metrics
  9. Loyalty
  10. Inverse Loyalty
  11. Loyalty Probabilities
  12. Inverse Loyalty Probabilities
  13. Experiments and Results
  14. Conclusion and Future Work
  15. Loyalty with all neighbors deleted for GNNs with self-loops
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Explainability techniques to determine the importance of the neighbors (Saliency Map simonyanDeepConvolutionalNetworks2014a in Figure \ref{['fig: saliency map']} and GNNExplainer ying2019gnnexplainer in Figure \ref{['fig: gnnexplainer']}). Both are computed for node classification in Cora dataset with a GCN with self-loops. The red node is the node being classified and the others are the neighbors used for the classification. The importances are in the range [0-1] with 1 being the maximum.
  • Figure 2: Loyalty and Inverse Loyalty of Saliency Map of GNNs with self-loops on Cora dataset
  • Figure 3: Loyalty and Inverse Loyalty Probabilities of Saliency Map of GNNs with self-loops on Cora dataset
  • Figure 4: Loyalty without self-loops