GNNExplainer for Graph Level Prediction Task

[amoschoomy]

    explainer = Explainer(
        model=model,
        algorithm=GNNExplainer(epochs=200),
        explanation_type='model',
        node_mask_type='attributes',
        model_config=dict(
            mode='binary_classification',
            task_level='graph',
            return_type='probs',
        ))

    sample_reprensatative_pred = []
    sample_reprensatative_label = []
    for i, sample in enumerate(loader):  # iterate over samples
        sample_scores = []
        sample_labels = []
        for data in sample:  # iterate over TCR sequences in each sample
            with torch.no_grad():
                out = model(data) 
                probabilities = torch.sigmoid(out)
                sample_scores.extend(probabilities[:, 1].tolist())
                sample_labels.extend(data.y.tolist())

How would one use an explainer to explain my graph level prediction task?

[rusty1s]

You can call the explainer similar to how you call the model, i.e.:

for i, sample in enumerate(loader):  # iterate over samples
        for data in sample:  # iterate over TCR sequences in each sample
                explanation = explainer(data.x, data.edge_index, ...)

Two things:

• The explainer can not operate on data objects. You need to change your model forward function to expect Tensor attributes instead: def forward(self, data) -> def forward(self, x, edge_index, ...)
• return_type should be log_probs in your case since the model does not seem to output probabilities (you have a sigmoid applied afterwards).