feature normalization

[mtzortzi]

Hi all,

I'm using PyG to make graph regression in a problem. I would like to ask you the following question.
I have my targets which are two values per graph, and one is smaller than the other, for example, I have one value in the range of 0.4 and the other value which I have to predict together is in the range of 8000.
So to make it work I have to normalize them somehow. My question is, should I normalize them together per row? or should I normalize each one of them separately and after that combine them to make my training?

Thank you in advance

[rusty1s]

I think it is best to normalize each target separately, e.g., via

mean, std = dataset.data.y.mean(dim=0, keepdim=True), dataset.data.y.std(dim=0, keepdim=True)

for batch in dataset:
    batch.y = (batch.y - mean) / std

[mtzortzi]

Great.
So, by separately we mean, calculating for the whole dataset (before splitting into train, valid, test) the mean+std for the target y1 and normalizing it, and then we do the same thing for target y2?

[rusty1s]

Indeed. If data.y is of shape [1, 2], then the above code will achieve this already.

[mtzortzi]

Thank you for your help :)

[Shiiivaa]

Great. So, by separately we mean, calculating for the whole dataset (before splitting into train, valid, test) the mean+std for the target y1 and normalizing it, and then we do the same thing for target y2?

You need to do normalization after splitting data into train and test. Because test data are centered and scaled according to the training data.