jma712 / gear Goto Github PK

Hi, May I ask why german dataset isn't used for calculating the counterfactual fairness subgraph and used for comparing the performance of node representation learning methods with respect to prediction and fairness?

Hi.

When reproducing your work, I came across a bug:
RuntimeError: scatter_mean() expected at most 5 argument(s) but received 6 argument(s)
I've fixed this and am writing this issue in case someone else come across the same bug.

Reason:

I found out this is caused by incompatible versions of PyG & torch_scatter.
Since you are using Pytorch 1.6.0, the only compatible versions of torch_scatter are 2.0.5 & 2.0.6 according to https://data.pyg.org/whl/, yet neither of these versions is compatible with PyG 1.3.0. To be specific, when using torch_scatter, PyG 1.3.0 calls function scatter_...(mean/add/...) with 6 arguments. However, in torch_scatter 2.0.5/2.0.6 the corresponding function is defined only with 5 arguments.

Solution:

(1) Using earlier versions of torch_scatter( and earlier versions of Pytorch since they depend on each other) or
(2) using later versions of PyG
will work. I change my PyG to 1.4.1 and the problem is fixed.

Thanks for the great work!