Why can GCNConv not handle negative edge weights?

[dinhtrang24]

Are there any reasons why graph convolutional operator GCNConv from the “Semi-supervised Classification with Graph Convolutional Networks” paper cannot handle negative edge weights? Is it because of mathematical complexity? Or are there any theoretical reasons/ assumptions, for example, if we violate them the model will fail? I have tried to find the answers in the paper, but it seems that the authors did not mention negative weight.

Thanks in advance!

[rusty1s]

The reason why GCNConv cannot support negative edge weights is due to the pow(-0.5) call on edge weights. If you do not perform such normalization, negative edge weights should work just fine.

[dinhtrang24]

Thanks a lot!