how to use pyG to do graph level prediction

[lingxuan9]

Hi!

Thanks for such amazing tool!

I have a use case:

There is a graph with 3000 edges, both nodes and edges have some features. When removing a edge from the graph, a score got. I would like to predict such score(related to number of nodes, shortest paths & other features).

How can I perform PyG to a graph level prediction? It seems a graph level prediction problem.

Thanks for any advice.

[wsad1]

You'll probably need to use some sort of pooling layer to pool node/edge features. See an example of using topk pool to make graph level predictions on a protein here. You can find all pooling layers in pyg here

[lingxuan9]

Thanks for your reply @wsad1 . Let me describe the question a bit more.

Assume that there is a graph with m nodes, n edges, every time when removing a or serval edge(s) from the graph, a defined score is calculated. Now I would like to predict the value of the score when removing a/some new edges from the graph.

My solution: convert this question into a graph level prediction question. After removing a/some edges from the graph, actually we can get/generate a new graph. So we can get a dataset contains a lot of graphs by random removing edges from the graph. Each graph is corresponding to a score -> predict label. Now the issue is predict such score for a given graph.

My questions:

• Does the solution good? Is there other solutions? Is there other similar use case?
• Can pyG perform such issue well?

Thanks!

[rusty1s]

Yes, I think this is exactly how you would approach this.

[lingxuan9]

Yes, I think this is exactly how you would approach this.

Thanks for your reply!

[lingxuan9]

Hi @rusty1s sorry for bothering you. Is there any graph level regression example I can follow? I found this graph classification example, but would like to have more infos. Thanks!

[rusty1s]

You can take a look here: https://github.com/pyg-team/pytorch_geometric/blob/master/examples/pna.py