I want to train a classifier on music data which contains a limited set of features which are all constrained in range:
name of note (7 possible values: C, D, ...)
Octave 0-7
Pitch alteration going from -2 to 2
duration as fraction of quarter notes - this is the only one why can vary a bit
I know it is best practice to encode and normalize after the split on the training set and then apply the learned operations on the test set. But does this apply here as well considering the limited range of these features and their well known values in music theory?
