Question about results on DBLP about hypergcn

[naganandy]

Hi,

Thanks for the interest in our work. The results reported in Table 4 are mean and standard deviation over different train-test splits with learning rate = 0.01. It is an interesting observation that the error can be (significantly) reduced by changing the learning rate!

[PlusRoss]

Thanks for your reply! I'm still confused about the train-test split. Do you mean that all the nodes are split into training or test set without any exception? If so, I notice that in the co-citation dataset, there exists a large portion of nodes (e.g. 40% in Cora) that do not appear in any hyperedge. Are they also included in the training or test set?

[naganandy]

yes, they are included without exception. GCN uses node's features (e.g. bag-of-words) even when the node does not appear in any hyperedge.

[PlusRoss]

Thanks for your answer!