How to take out the important subgraphs of the whole graph?

[yung1231]

Hello, I built a model to classify the whole graph and I tried to use explainer.explain_graph to get those parts that are important.

Then I got two returns node_feat_mask、edge_mask.

Then I use explainer.visualize_subgraph to visualize the result but I see that the output is still the whole graph, not the subgraph.

Is there a way to show only the important subgraphs instead of the whole graph? Because the number of nodes and edges of the whole graph is too large, which makes it impossible to observe.

---

Then there is one more question, each time I run the result, the values of node_feat_mask and edge_mask are different, is this normal?

---

explainer = GNNExplainer(model, epochs=3)    # only be executed once

# The following execution, each time the result is different
node_feat_mask, edge_mask = explainer.explain_graph(data.x, data.edge_index) 
ax, G = explainer.visualize_subgraph(None, edge_index, edge_mask, y=data.y, threshold=0.5)  # mask less than

Thanks.

[rusty1s]

One solution would be to manually shrink your graph to only show important edges and nodes, e.g. via
subgraph(node_feat_mask > 0.5, edge_index). You can visualize the result afterwards.

GNNExplainer has indeed some variance in its result, but you may be able to fine-tune its hyperparameters a bit to get more stable results.

[rusty1s]

Here is an example.

[yung1231]

No no, what I want is to classify the whole graph(is graph classification, not node classification), and then only show the important subgraph (Instead of showing the whole graph, only the important subgraph is displayed).
I use explainer.explain_graph and then use explainer.visualize_subgraph but the explainer.visualize_subgraph result shows the whole graph, not just the important subgraphs.

I want to show only the important subgraphs because the number of nodes on the graph is so large that it is completely unobservable if the whole graph is displayed.

[rusty1s]

Yes, this brings us back to my comment earlier. You need to filter the "unimportant nodes" and edges on your own. For example, if feat_mask_type = "scalar", you can do

node_mask = node_feat_mask > 0.5
edge_mask = edge_mask > 0.5

edge_index = edge_index[:, edge_mask]  # Filter important edges
edge_index, _ = subgraph(node_mask, edge_index, relabel_nodes=True)  # Filter important nodes

[yung1231]

Ummm, It looks like something, but some use edge_index, _ = subgraph(node_mask, edge_index) to run the result, the edge_index will be empty, both large and small graphs have this situation.

[rusty1s]

You might want to adjust the thresholds to a lower value.