Explainability with SAGPooling layer

[giuseppeantoniomotisi]

Hello everyone,

I’m currently working on explainability in Graph Neural Networks (GNNs) and facing some challenges.
My model architecture consists of:

• (SAGEConv) × 2 + SAGPooling, followed by
• global_max_pool concatenated with global_mean_pool, leading to
  A classification head for a whole-graph binary classification task.
  I am able to extract relevant subgraphs using pooled nodes and their corresponding attention scores. However, I am looking for a way to validate these explanations, potentially using GNNExplainer. Unfortunately, I’ve read that GNNExplainer doesn’t work well with dynamic pooling layers like SAGPooling (which involves nested TopKPooling).
  I also tried PGExplainer, but the results were not promising. I observed:
• Fidelity values of {(0,0)}
• A loss that doesn’t decrease during training (stucked at 1.344... also changing LR and number of epochs)

Does anyone have suggestions on how to explain this model effectively?
Any help would be greatly appreciated!

Thanks in advance!