Node feature learning

[levtelyatnikov]

Hi, thank you for providing the code! I have one question regarding the node features, in the paper I got the impression that the node features are initialized with random walk, while in the code of train.py file, line 392 I can see that you also learned the node features:
initembedder = Wrap_Embedding(data.numnodes, args.input_vdim, scale_grad_by_freq=False, padding_idx=0, sparse=False)

I cannot find any mention of learning the node features in the paper, can you please provide some insights regarding this? Have you tried to make them unlearnable, how do the results change then?

I am wondering because the datasets are quite big, hence making the node embeddings learnable adds a lot of computational complexity.

Thank you once more for the code!!

[young917]

Hello, and thank you for your feedback!

You are correct that the node features are indeed learnable. However, we would like to note that the embedding layer is initialized through random walk, as seen in the code at line 393 of the train.py file.

Our decision to initialize node features in this manner was based on the implementation provided in the repository:

EdgeDependentNodeLabel/train.py

Line 393 in 5f2d6b3

initembedder.weight = nn.Parameter(A)

. While we haven't specifically experimented with making the node features unlearnable, we believe it's a valuable idea worth exploring.
Furthermore, it's worth mentioning that this approach is consistently applied across all baseline approaches. So we don't consider it to be the factor determining our performance. However, we appreciate your suggestion and see the value in experimenting with making the node features unlearnable.

Thank you once again!