search

snap-stanford / CAW

MIT License
73 stars 19 forks source link

Features of datasets #3

Closed Cantoria closed 3 years ago

Cantoria commented 3 years ago

Hi dear authors, I have noticed that the processed file contains node features and edge features for all datasets. As far as I know, the original ENRON/UCI/social-evolution datasets do not contain any node/edge features. I have checked your paper, but the paper doesn't mention any details for initializing node or edge features. Can you describe the details of initializing node or edge features in these .npy files? Thanks!

lipan00123 commented 3 years ago

I guess we initialize them as constant features to guarantee the model is inductive.

Cantoria commented 3 years ago

I guess we initialize them as constant features to guarantee the model is inductive.

How to initialize features of three datasets? randomly initialize?

lipan00123 commented 3 years ago

We used "constant" features as initial features. We have tried random initialization, which typically gives better transductive learning performance, worse inductive learning performance and the model converges slow.

Cantoria commented 3 years ago

We used "constant" features as initial features. We have tried random initialization, which typically gives better transductive learning performance, worse inductive learning performance and the model converges slow.

Thanks for your reply. You would better mention this in your paper.

lipan00123 commented 3 years ago

One thing should be emphasized that. Our method does not need node features at all as we may capture the structural dynamics.