Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CF-GNNExplainer: Counterfactual Explanations for Graph Neural Networks

Ana Lucic, Maartje ter Hoeve, Gabriele Tolomei, Maarten de Rijke, Fabrizio Silvestri

TL;DR

CF-GNNExplainer introduces a graph-specific counterfactual explanation method that perturbatively prunes edges in a node's local subgraph to flip predictions. By framing CF generation as an adjacency perturbation problem and optimizing a joint loss, it yields minimal, sparse, and accurate counterfactuals across three synthetic GNN explanation datasets. The approach outperforms baselines and a standard GNNExplainability method in the CF setting, while highlighting the importance of evaluation protocols and societal considerations for CF explanations. Limitations include edge deletions only and node-classification focus, with planned extensions to feature perturbations and graph-level tasks.

Abstract

Given the increasing promise of graph neural networks (GNNs) in real-world applications, several methods have been developed for explaining their predictions. Existing methods for interpreting predictions from GNNs have primarily focused on generating subgraphs that are especially relevant for a particular prediction. However, such methods are not counterfactual (CF) in nature: given a prediction, we want to understand how the prediction can be changed in order to achieve an alternative outcome. In this work, we propose a method for generating CF explanations for GNNs: the minimal perturbation to the input (graph) data such that the prediction changes. Using only edge deletions, we find that our method, CF-GNNExplainer, can generate CF explanations for the majority of instances across three widely used datasets for GNN explanations, while removing less than 3 edges on average, with at least 94\% accuracy. This indicates that CF-GNNExplainer primarily removes edges that are crucial for the original predictions, resulting in minimal CF explanations.

CF-GNNExplainer: Counterfactual Explanations for Graph Neural Networks

TL;DR

CF-GNNExplainer introduces a graph-specific counterfactual explanation method that perturbatively prunes edges in a node's local subgraph to flip predictions. By framing CF generation as an adjacency perturbation problem and optimizing a joint loss, it yields minimal, sparse, and accurate counterfactuals across three synthetic GNN explanation datasets. The approach outperforms baselines and a standard GNNExplainability method in the CF setting, while highlighting the importance of evaluation protocols and societal considerations for CF explanations. Limitations include edge deletions only and node-classification focus, with planned extensions to feature perturbations and graph-level tasks.

Abstract

Given the increasing promise of graph neural networks (GNNs) in real-world applications, several methods have been developed for explaining their predictions. Existing methods for interpreting predictions from GNNs have primarily focused on generating subgraphs that are especially relevant for a particular prediction. However, such methods are not counterfactual (CF) in nature: given a prediction, we want to understand how the prediction can be changed in order to achieve an alternative outcome. In this work, we propose a method for generating CF explanations for GNNs: the minimal perturbation to the input (graph) data such that the prediction changes. Using only edge deletions, we find that our method, CF-GNNExplainer, can generate CF explanations for the majority of instances across three widely used datasets for GNN explanations, while removing less than 3 edges on average, with at least 94\% accuracy. This indicates that CF-GNNExplainer primarily removes edges that are crucial for the original predictions, resulting in minimal CF explanations.

Paper Structure

This paper contains 26 sections, 5 equations, 2 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. GNN Explainability
  4. Counterfactual Explanations
  5. Adversarial Attacks
  6. BACKGROUND
  7. Graph Neural Networks
  8. Matrix Sparsification
  9. PROBLEM FORMULATION
  10. METHOD: CF-GNNExplainer
  11. Adjacency Matrix Perturbation
  12. Counterfactual Generating Model
  13. Loss Function Optimization
  14. CF-GNNExplainer
  15. Complexity
  16. ...and 11 more sections

Figures (2)

  • Figure 1: Intuition of counterfactual example generation by CF-GNNExplainer.
  • Figure 2: Histograms showing the proportion of CF examples that have a certain explanation size from random. Note the $x$-axis for ba-shapes goes up to 1500. Left: tree-cycles, Middle: tree-grid, Right: ba-shapes.