Does LaplacianPositional encoding matrix correspond to the same order as the node feature matrix?

[amjass12]

Hi all!

I am running Laplacianeigenvector positional encoding on my torch geometric dataset, however, I am unsure about whether the PE matrix returned is in the same order as in the node features in the torch dataset. Please can somebod confirm? many thanks.

code:

from torch_geometric.transforms import AddLaplacianEigenvectorPE
pos_enc = AddLaplacianEigenvectorPE(k=50)
pos_enc(graph_dataset)

now 'laplacian_encoding_pe' matrix is added as expected, but is the order of the matrix the same as the node features (data.x)?

so is PE[0] the positional encoding for node 0 (as in data.x[0]) and so on

Many thanks!!

[rusty1s]

Yes, this should work. I added a test to confirm: #8087

[amjass12]

Hi @rusty1s ,

thank you very much for confirming! would it be possible to implement an option for padding if the selected k has more dimensions than eigenvectors?

thank you!

[rusty1s]

Yes, that is possible. Do you wanna contribute this?

[amjass12]

for sure, will start working on this as soon as i have time! thanks again!