test set normalization based on training set

[snurkabill]

[mihaelacr]

Hi! Thanks for the contribution!

Your commit and pull request say 'test set normalization based on training set', but I do not see the gaussianNormalization function that takes two parameters ever used.

Also overloading does not work like this in python: you cannot define two functions wit the same name, you have to use default arguments, please see a discussion here.

Did you try running the with your change? Do you see any improvements?

[snurkabill]

Hi,

I must admit that I didn't run those changes. I have some local changes that works and I've just tried to put it together.

first of all I just wanted discuss those changes , I will make proper PR later. my point right here is, that test set's normalization should be based on atributes gained on training set.

normalization itself is done in normalizeData function

[snurkabill]

Btw, I suspect that all of my PR's won't work for the first time. I really want to know your opinion :)

[snurkabill]

ok, scale() is renamed as it was before.

Motivation for scaling testing data based on training set:

• when data is online, we can't normalize testing set on it's parameters because we don't have all data yet, we need to use already measured params
• model is based on some noramlized data, we should also use that normalization for testing data... if testing data are somehow different (shifted atc.) we have wrongly sampled training set.