Correct way of normalizing the node features

[SaffronWolf]

I am working on graph classification problem and was wondering how to normalize node features. Are they supposed to be normalized individually or for entire graph at a time? Also, how do you ensure the same behavior at inference?

[rusty1s]

You should calculate mean and standard deviation of node features across all nodes in all training graphs:

mean = train_dataset.data.x.mean(dim=0, keepdim=True)
std = train_dataset.data.x.std(dim=0, keepdim=True)

You can then apply normalization within every graph:

data.x = (data.x - mean) / std

[SaffronWolf]

Thank you so much. Doing this gave me a freaking 50% jump in accuracy !!!!

[rusty1s]

Awesome :)