The function graph_split() seems to contradict the inductive scenarios. about graphless-neural-networks

[ShichangZh]

Thank you for your question.

The reason why idx_test_ind and obs_idx_train can have the same element is because they correspond to two different indexing systems.

It is a bit complicated to create a proper inductive setting, though we try to make variable meanings clear in the function comments.

Indices start with 'obs_' correspond to the node indices within the observed subgraph,
where as indices start directly with 'idx_' correspond to the node indices in the original graph

In other words, ind_ covers [0, N], and obs_idx covers [0, N_{obs}], where the 0 in these two indexing systems corresponds to different nodes.

Let me explain it in more detail.

If you look at the following line in the run_inductive function, you will see that idx_obs is meant to index the original graph and construct the observed graph.

graphless-neural-networks/train_and_eval.py

Line 324 in a24d570

obs_g = g.subgraph(idx_obs)

This subgraph function will assign each node a new index. Later on, when creating a nodeloader in the following line, we pass obs_idx_train to the loader function to index the observed graph (subgraph).

graphless-neural-networks/train_and_eval.py

Line 336 in a24d570

obs_dataloader = dgl.dataloading.NodeDataLoader(

In other words, obs_idx_train corresponds to the new indices created by the subgraph function.

Sorry for the confusion. I hope this helps. Please let me know if you have other questions.

[Leon-0602]

Thanks a lot for your clear explanations and patient help. I believe that GLNN and your future work will be very impactful. :)