Data split used in Table 1

[zknus]

Hi tk-rusch,

I am confused about how you conduct the random data split in Table 1 of Graphcon paper.
From the description and the test ACC of the baseline model, table 1 follows the same data split ratio in [1], which is 20 per class for training, 30 per class for val and the rest of data for test.
However, from the code,

GraphCON/src/homophilic_graphs/data.py

Lines 74 to 78 in 3326c48

	if use_lcc or not train_mask_exists:
	dataset.data = set_train_val_test_split(
	12345,
	dataset.data,
	num_development=5000 if ds == "CoauthorCS" else 1500)

It seems that Graphcon only uses 20 per class for training and the val and test ratio is different with the paper [1].
Can you explain the train/val/test ratio you used in table 1? Thanks very much!

[1] Shchur O, Mumme M, Bojchevski A, et al. Pitfalls of graph neural network evaluation[J]. arXiv preprint arXiv:1811.05868, 2018.