Rex Ying

I think yes you need to try both to see which one is better. Another advantage is that GraphSAGE can capture local structure similar to WL test. And it is...

You will need the features for all nodes as well. When splitting train/test, you can split the edges, not nodes. So 20% edges as test; 80% edges as training.

Yes the answer is correct. In addition, you can also add the link pred as auxiliary loss for supervised learning.

this is great! thanks. For sparse features, this would be useful. In general, very large feature arrays without sparsity are hard to deal with, especially since the minibatch algorithm requires...

(1) Here we are defining length by the size of the (sampled) neighborhood. (2) neigh_h is the neighborhood embedding. Here gather is just to extract the embeddings corresponding to the...

Hi yes let me make an update

Yes there will be stochasticity. With ~10 and 25 neighbors at each layer we do not observe too much variability. Also sample can be performed at training time only. At...

For raw data, http://snap.stanford.edu/ohmnet/ has the graph structure; http://software.broadinstitute.org/gsea/msigdb/collections.jsp has the feature and label information. c1, c3, c7 are the feature sets; GO is the label set.

yes nx has made several updates. i'll be happy to merge in fix that works with new networkx version and update requirements.

Hi, to run on own dataset, one should first train the GNN for the prediction task. You can use the base implementation in the repo, or replace with your own....