The comparisons are unfair due to the use of feature expansion.

[cszachary]

Hi,

I find the comparisons with baselines are unfair in unsupervised TU, since existing baselines for plain datasets only use one dimensional attribute (i.e., the attribute of all nodes is set as 1). Please see here (https://github.com/Shen-Lab/GraphCL/blob/1d43f79d7f33f8133f9d4b4b8254d8aaeb09a615/unsupervised_TU/gin.py#L47). Without the using of feature expansion, the performance of AutoGCL drops significantly.

[Somedaywilldo]

Hi, would you mind sharing your settings, are you using the default settings for both models, I'll double check.

[cszachary]

Hi,

I use the default setting and run un_exp.sh. For attributed graph, I use their original features without feature expansion. For plain graph, I set their features as 1. The performance drops in difference scales on all the datasets (i.e., from 1% ~ 4%).