How can I use GNNExplainer and visualize the subgraph?

[masteryqq]

Hi,

I tried the tutorials for XAI, however, although PyG has been renewed, the tutorials of GNN has not been renewed yet. The example offered in this blog post has been deleted. https://medium.com/@pytorch_geometric/graph-machine-learning-explainability-with-pyg-ff13cffc23c2

This is the code I used to construct the GNN on MUTAG dataset. https://colab.research.google.com/drive/1fLJbFPz0yMCQg81DdCP5I8jXw9LoggKO?usp=sharing

Then I applied code comes from the blog post:

from torch_geometric.explain import Explainer, GNNExplainer
# Initialize explainer
explainer = Explainer(
    model=model,
    algorithm=GNNExplainer(epochs=200),
    explanation_type='model',
    node_mask_type='attributes',
    edge_mask_type='object',
    model_config=dict(
        mode='multiclass_classification',
        task_level='node',
        return_type='log_probs',
    ),
)

It couldn't offer explanation for me. Could you please offer examples of GNNExplainer for graph classification problem? Besides, could you please let me know how to visualize the subgraph of explanation?

Thank you!

-David

[wsad1]

Check out the example here. You need to call visualize_graph to visualize the explanations.

[wsad1]

You could pass just one batch to the explainer to get more interpretable explanations.
Other than visualizing the fimportance graph. You could visualize the feature importance, or simply view the available_explanations. See here

[masteryqq]

Hi,

Thank you for your reply! When I had a look at the examples you offered: GNNExplainer on BAShapes, I have questions.

1. Why do we need to use node index?

I thought it is a graph classification problem, so the code should be similar to MUTAG graph classification problem.

### Initialize explainer
explainer_pyg = Explainer(
    model=model,
    algorithm=GNNExplainer(epochs=200),
    explanation_type='model',
    node_mask_type='attributes',
    edge_mask_type='object',
    model_config=dict(
        mode='binary_classification',
        task_level='node',
        return_type='raw',
    ),
)

explanation = explainer_pyg(
    x=data.x,
    edge_index=data.edge_index,
    batch=data.batch,
    target=data.y
)

Then I could use explanation.visualize_graph to plot the subgraph.

2. Why do we need to train the explainer?

Since GNNExplainer is an instance-level explainer, we could directly generate explanation for a graph without training process. Maybe according to this step you would like to calculate AUC?

Thanks！

-David

[rusty1s]

1. This is a node-level classification task, and so we iterate over all node indices here for which we want to obtain an explanation.
2. GNNExplainer will train each instance separately. Maybe I misunderstood your question though.

[masteryqq]

Hi,

Thank you for your reply! I was confused since this step is not necessary for graph classification task, but I understand this step now.

-David

[sumone-compbio]

@masteryqq hi, did you figure out how to represent molecules with GNNExplainer results? There is an option to highlight molecules' substructures in rdkit (https://www.rdkit.org/docs/GettingStartedInPython.html) but you have to provide the string for the substructure. Highlighting nodes/edges in gradient based on importance scores that we get from GNNExplainer would be nice.