I want to use an interpretability algorithm for heterogeneous graphs（like from torch_geometric.explain import Explainer, CaptumExplainer）, how should I do it

[GXFay]

I want to use an interpretability algorithm for heterogeneous graphs（like from torch_geometric.explain import Explainer, CaptumExplainer）, how should I do it

[rusty1s]

Can you clarify on what you mean by "interpretability"? Can't you use use CaptumExplainer (which supports heterogeneous graphs)?

[GXFay]

when I use CaptumExplainer, it told me (RuntimeError: One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.)

[rusty1s]

Mh, I guess this happens if you have node types/edge types that are not reachable within message passing. @wsad1 Is there a way to fix this?

[GXFay]

Thanks for u answer. I try to use to_captum_model, but it still report an error:

[rusty1s]

How do node_types and x_dict.keys() look like?

[yfeng77]

Hi, I got the same error. Did you find a solution? And when I run

hetero_explanation = explainer(
    data.x_dict,
    data.edge_index_dict,
    index=torch.tensor([1, 3]),
) 

again without changing anything, then I got

'IntegratedGradients' object is not callable.

error.

[yfeng77]

Do you have a reproducible example? We added an example of using CaptumExplainer on heterogeneous graphs in #7096.

Hi. Thank you for your response. I am working on the Explaining node classification on a heterogeneous graph that I found on the PyG tutorial . For the hetero_data, I tried the IMDB dataset and OGB_MAG dataset. For model, I tried HeteroGNN, HGT and HAN. But none of them work. When I used the IMDB dataset and the HAN model, I got the same error as the author got. Thank you!

[rusty1s]

Do you have a small example to share so that I am able to reproduce?

[yfeng77]

Do you have a small example to share so that I am able to reproduce?

Yes. Hope this example will help. Thank you so much.
Untitled (1).pdf

[rusty1s]

Sorry for super late reply. I assume this is because our heterogeneous explainer is designed with to_hetero or HeteroConv in mind, and so it fails to explain heterogeneous graphs where there does not exist separate GNN layers for each edge type.
This is temporarily fixable by:

explainer = Explainer(
    model=model,  # It is assumed that model outputs a single tensor.
    algorithm=CaptumExplainer('IntegratedGradients'),
    explanation_type='model',
    node_mask_type='attributes',
    edge_mask_type=None,
    model_config=dict(
        mode='multiclass_classification',
        task_level="node",
        return_type='probs',  # Model returns probabilities.
    ),
)

but this is obviously not a great workaround.

[yfeng77]

Thank you! I may try another GNN model.