Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

Hyeoncheol Cho, Youngrock Oh, Eunjoo Jeon

TL;DR

This work tackles explainability for graph neural networks by focusing on node-classification explanations. It introduces SEEN, a post hoc method that sharpens a target explanation by aggregating auxiliary explanations from nearby assistant nodes within the graph, weighting them by their importance in the target explanation via an exponential decay governed by α and β: $\bar{S}(v_t) = S(v_t) + \alpha \sum_{r=1}^{|V_a|} \beta^{r-1} S(v^{(r)})$. The approach is model- and graph-agnostic, requiring no graph modification, and is compatible with multiple explainability techniques such as SA, Grad*Input, and GradCAM. Experiments on four synthetic datasets show improvements in explanation accuracy up to 12.71%, with analysis highlighting the importance of ranking-based weighting and parameter sensitivity. SEEN thus offers a practical, post-hoc mechanism to enhance GNN explanation reliability by leveraging local neighborhood information without retraining or data changes.

Abstract

Explaining the foundations for predictions obtained from graph neural networks (GNNs) is critical for credible use of GNN models for real-world problems. Owing to the rapid growth of GNN applications, recent progress in explaining predictions from GNNs, such as sensitivity analysis, perturbation methods, and attribution methods, showed great opportunities and possibilities for explaining GNN predictions. In this study, we propose a method to improve the explanation quality of node classification tasks that can be applied in a post hoc manner through aggregation of auxiliary explanations from important neighboring nodes, named SEEN. Applying SEEN does not require modification of a graph and can be used with diverse explainability techniques due to its independent mechanism. Experiments on matching motif-participating nodes from a given graph show great improvement in explanation accuracy of up to 12.71% and demonstrate the correlation between the auxiliary explanations and the enhanced explanation accuracy through leveraging their contributions. SEEN provides a simple but effective method to enhance the explanation quality of GNN model outputs, and this method is applicable in combination with most explainability techniques.

SEEN: Sharpening Explanations for Graph Neural Networks using Explanations from Neighborhoods

TL;DR

This work tackles explainability for graph neural networks by focusing on node-classification explanations. It introduces SEEN, a post hoc method that sharpens a target explanation by aggregating auxiliary explanations from nearby assistant nodes within the graph, weighting them by their importance in the target explanation via an exponential decay governed by α and β: . The approach is model- and graph-agnostic, requiring no graph modification, and is compatible with multiple explainability techniques such as SA, Grad*Input, and GradCAM. Experiments on four synthetic datasets show improvements in explanation accuracy up to 12.71%, with analysis highlighting the importance of ranking-based weighting and parameter sensitivity. SEEN thus offers a practical, post-hoc mechanism to enhance GNN explanation reliability by leveraging local neighborhood information without retraining or data changes.

Abstract

Explaining the foundations for predictions obtained from graph neural networks (GNNs) is critical for credible use of GNN models for real-world problems. Owing to the rapid growth of GNN applications, recent progress in explaining predictions from GNNs, such as sensitivity analysis, perturbation methods, and attribution methods, showed great opportunities and possibilities for explaining GNN predictions. In this study, we propose a method to improve the explanation quality of node classification tasks that can be applied in a post hoc manner through aggregation of auxiliary explanations from important neighboring nodes, named SEEN. Applying SEEN does not require modification of a graph and can be used with diverse explainability techniques due to its independent mechanism. Experiments on matching motif-participating nodes from a given graph show great improvement in explanation accuracy of up to 12.71% and demonstrate the correlation between the auxiliary explanations and the enhanced explanation accuracy through leveraging their contributions. SEEN provides a simple but effective method to enhance the explanation quality of GNN model outputs, and this method is applicable in combination with most explainability techniques.

Paper Structure

This paper contains 31 sections, 3 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Graph neural networks
  4. Explainability methods for GNN
  5. Preliminaries
  6. Graph neural networks
  7. Explaining GNN predictions
  8. $\textsc{Seen}$: Sharpening Explanations using Explanations from Neighborhoods
  9. Auxiliary explanation acquisition
  10. Explanation aggregation
  11. Motivation
  12. Experiments
  13. Datasets
  14. Model
  15. Explainability methods
  16. ...and 16 more sections

Figures (3)

  • Figure 1: Schematic for sharpening the explanation with $\textsc{Seen}$ on node classification. With the input graph, trained model and explanation target, $\textsc{Seen}$ performs explanation sharpening through selecting assistant nodes, generating auxiliary explanations from the assistant nodes, and aggregating the target and auxiliary explanations.
  • Figure 2: Effects of aggregation coefficients on explanation accuracy for BA-Community and Tree-Grid datasets. The results for the remaining combinations are listed in the Appendix.
  • Figure 3: Effects of aggregation coefficients on explanation accuracy for BA-Shapes, BA-Community, Tree-Cycles, and Tree-Grid datasets.