GNNExplainer with PyGOD ?

[Srivathsan-Srinivas]

Hi,
I just began using PyGOD (to find a contextual-anomalous node). It works well and is able to find the anomalous node (the particular row in the input dataframe).

Now, I want to find the reasons why this particular row is anomalous. each node/row has about 35 features and I want to know which of these features contributed the most to the anomaly. So, I tried GNNExplainer like below

explainer = Explainer(
    model=trained_ad_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',  # Model returns log probabilities.
    ),
)

# Generate explanation for the node at index `10`:
explanation = explainer(graph_data.x, graph_data.edge_index, index=10)
print(explanation.edge_mask)
print(explanation.node_mask)

Traceback (most recent call last):
  File "/Users/sri/exps/graphs/src/main.py", line 696, in <module>
    main(get_args())
  File "/Users/sri/exps/graphs/src/main.py", line 544, in main
    explanation = explainer(graph_data.x, graph_data.edge_index, index=10)
  File "/Users/sri/.pyenv/versions/ve_3.10.13/lib/python3.10/site-packages/torch_geometric/explain/explainer.py", line 195, in __call__
    prediction = self.get_prediction(x, edge_index, **kwargs)
  File "/Users/sri/.pyenv/versions/ve_3.10.13/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/Users/sri/.pyenv/versions/ve_3.10.13/lib/python3.10/site-packages/torch_geometric/explain/explainer.py", line 111, in get_prediction
    training = self.model.training
AttributeError: 'DOMINANT' object has no attribute 'training'
make: *** [run] Error 1

Is there a way I can use GNNExplainer with PyGOD's algorithms (eg: DOMINANT)?

[kayzliu]

Thanks for your interest to PyGOD. Using GNNExplainer with PyGOD may be feasible, but I am not sure which implementation exactly you are using and how you define trained_ad_model. If trained_ad_model is an instance of pygod.detector.DOMINANT, feeding the explainer trained_ad_model.model instead of trained_ad_model may help.

[Srivathsan-Srinivas]

Yes, trained_ad_model is an instance of DOMINANT. When I print the type, it prints it as type of trained anomaly detector:<class 'pygod.detector.dominant.DOMINANT'> (slightly different from what you suggested to check for). Add, I can't add .model to this instance: It complains AttributeError: 'tuple' object has no attribute 'model'.

I am following the instructions from https://pytorch-geometric.readthedocs.io/en/latest/tutorial/explain.html and from https://medium.com/@pytorch_geometric/graph-machine-learning-explainability-with-pyg-ff13cffc23c2.