Are there any relevant existing networks or examples that make use of edge characteristics?

[boyin8]

Hi, I want to use both edge attributes and node attributes to do some link prediction tasks. Does PyG have a relevant example?

Many thanks!

[rusty1s]

I don't think we have an example for that, but it should be pretty straightforward to build from an existing link prediction example and just enhance the GNN with edge feature support. Let me know if you need any more guidance :)

[boyin8]

Hi, if I want to add my own defined edge weights to the AddMetaPaths, how to do that? I noticed that if I set the weighted parameter to True (The weight of each metapath edge is computed as the number of metapaths from the start to the end of the metapath edge), which is not what I want. Thanks.

Sorry, In edge_weight = data[edge_type].get('edge_weight', None), does this mean that the weights I defined are already being used?

[rusty1s]

There is a weighted option in AddMetaPaths which you need to turn on.

[boyin8]

Thank you for your advice. However, I've come across a problem regarding whether different network structures can be shared. Specifically, the data I generate randomly each time has the same number of nodes as well as the same type. The only thing that differs from one dataset to another is their network structure, i.e. the edge_index. For example, I currently have four sets of training data, A,B,C,D. Can I take the trained model from A and test it on data B? (In practice it seems to be possible and the model returns predictions in PyG).

The reason I need this operation is also because each time my nodes are generated randomly. Or do you have any other recommended methods in PyG?

[rusty1s]

Yes, a GNN can operate on unseen network structures. For example, you can train on A and B, validate on C and test on D.

[rusty1s]

Take a look at https://colab.research.google.com/drive/1xpzn1Nvai1ygd_P5Yambc_oe4VBPK_ZT?usp=sharing