Using GraphSage to do unsupervised node embeddings for node regression

[lfangyu09]

Hi,

Thank you so much for this package. I want to use GraphSage to do unsupervised node embeddings for node regression. May I ask three questions:

1. The example (examples/graph_sage_unsup.py) seems to use for classification loss = F.binary_cross_entropy_with_logits(pred, batch.edge_label). How can I change the loss function? Could it be changed to loss = F.mse_loss(pred, batch.edge_label)?
2. After training GraphSage, how can I use it to get node embedding? By model(data.x, data.edge_index)?
3. Is test() necessary? As I use my own customed data, how to set train_mask, val_mask, and test_mask?

Thank you so much!

[rusty1s]

1. How do you want to do unsupervised learning for node regression? If you wanna use mse_loss, you need ground-truth targets.
2. Yes :)
3. No, it's not necessary. It's just implemented here to test the quality of node embeddings.

[lfangyu09]

Thanks a lot for your quick reply.

1. I want to use GraphSage to get the positional encoding, similar to AddRandomWalkPE or AddLaplacianEigenvectorPE. Could I use this example (examples/graph_sage_unsup.py) to get the positional encoding?
2. May I ask what is the model output of this example (examples/graph_sage_unsup.py)?

Thanks.

[rusty1s]

Yes, you can basically use graph_sage_unsup.py as it is to get pre-trained node embeddings, which you can then use to perform a final regression downstream task. The model output here are basically pre-trained node embeddings.