04_interpretability

Interpreting a GNN Model

This experiment demonstrates how to interpret the behavior of a trained graph neural network (GNN) using the built-in interpreter GNNExplainer. GNNExplainer allows you to determine which nodes, features, and connections influenced the model's decision making. The GNN-AID framework supports different interpreters. This example uses GNNExplainer. You can easily change the interpreter in the configuration if you want.

---

Folder contents

• gnn_explainer_example.py — script for training GNN.
• run_example.sh — script for running the experiment.
• imgs/ — folder with images.
• README.md — description of the experiment.

---

Quick start

Run:

  bash run_example.sh

The script trains a GNN model on the Cora dataset and then runs the GNNExplainer interpreter to get a local interpretation for a particular node.

---

Interpretation results

The results of interpretation are automatically saved in the ./explanations folder. If you are using the web interface, you will see the resulting interpretation in a similar form:

_{Pic. 1. The result of the GNNExplainer algorithm on the Cora dataset (interpretation of the model's prediction for vertex 691).}

The algorithm highlighted influential edges and vertices in the graph. On the right side of the image, you can see the importance indicator for the highlighted vertices and edges.