Use this model explain the graph-level GNN model?

[skaudrey]

Hi, I have read and tried the codes, and now it can work on my graph-level GNN model in multi-instance case. The input graph in my project is a spatiotemporal graph, which connects the subgraph in each frame together as a whole graph.

Then I realize the difference between my project and the applications of GNNExplainer. The labels used in GNNExplainer are the node labels, while I am more interested to explain the relation between the GNN model and the whole graph’s label.

Is it possible that if I change the label in GNNExplainer to the graph label, can GNNExplainer then generate a meaningful explanation? Considering the mutual information in GNNExplainer, I think it can work, but I am not confident enough.

Could someone please give me some advice? This will be very helpful.

@RexYing @hnaik

[asarigun]

Hi @skaudrey,

I'm also trying to use GNNExplainer for the graph regression task. However, instead of finding existing edges in the graph, GNNExplainer generates new edges. For example, I have a subgraph with 238 nodes and 154 edges that GNNExplainer finds; just nine edges exist in the accurate graph.

Did you ever experience the same issue? I would be happy to hear your suggestions! @RexYing @hnaik

Best,

Ahmet