LSTM for prediction of next location step - help with standardization

Question

I have a few questions regarding the topic and I hope someone might have experience with any of them.

What I am trying to do is train an LSTM network, whose input is a sequence of N steps in a XYZ space (i.e 3 features over N point per sample, each point is part of some coordinate in space) and i want to predict the next point in the same XYZ space.

Note: Not all samples are from the same DB, meaning the XYZ space varies between some samples/

My questions are:

1. Say I want to use min-max scaling, do I scale across ALL samples at once or scale per XYZ space? say i have 100 samples from XYZ_A and 50 from XYZ_B, do i take my min/max from 150 point of both space or not?
2. should I also scale my output labels? and if so should I use them in the initial scale or should I only scale on my train inputs and use that scaler on my outputs?
3. When I want to make a prediction after training, should I scale the data according to my training data?

OR since my features are relatively from the same domain should I skip scaling all together?

Answer 1

Say I want to use min-max scaling, do I scale across ALL samples at once or scale per XYZ space? say i have 100 samples from XYZ_A and 50 from XYZ_B, do i take my min/max from 150 point of both space or not?

Even though your samples come from different databases, they represent the same physical quantity, i.e. 3D coordinates. Since the units of measurement are the same for the samples of both databases, you should apply the scaling over all 150 samples, not separately.

should I also scale my output labels? and if so should I use them in the initial scale or should I only scale on my train inputs and use that scaler on my outputs?

I wouldn't use the word labels, since you are predicting the next XYZ coordinates, therefore it is a regression problem, not classification. However, you should also apply the scaling function on those values too, since they have the same units of measurement as the input.

When I want to make a prediction after training, should I scale the data according to my training data?

Exactly, usually you fit a scaling function to your training dataset and you apply the same (already fit) function on the testing dataset.