PMStandardScaler fails when scale = 0

[AtharvaKhare]

A relevant example:

x := PMMatrix rows: #(#(1 0) #(1 0) #(1 1) #(1 1)).
scaler := PMStandardizationScaler new.
scaler fit: x.
scaler scale. "0"
scaler transform: x. "ZeroDivideError"

Possible fix: add a "handleZeroScale" method which changes scale to 1 when it is 0.
Question: Can this be method be added in PMDataTransformer? If we add more scalers in future, all would need this message.

[SergeStinckwich]

Why we should modify the scale to 1 when this is 0 ? 😄

[SergeStinckwich]

we should have a look how this is done in scikit-learn : https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.StandardScaler.html

[AtharvaKhare]

We are doing (element - mean) / scale. When scale=0, we can't divide by zero. Since all elements are same, the element-mean will be 0 as it should be. So any number in the denominator will be fine. 1 specifically because any number divided by it is itself, but again, any number should be ok.

sklearn is also setting scale to 1.

[SergeStinckwich]

ok great if they use the same trick. I find the Python code here: https://github.com/scikit-learn/scikit-learn/blob/7813f7efb/sklearn/preprocessing/data.py#L62

[SergeStinckwich]

Can you propose a fix, should be easy ?

[AtharvaKhare]

One of the fix is to modify this method:

PolyMath/src/Math-PrincipalComponentAnalysis/PMStandardizationScaler.class.st

Lines 24 to 26 in aaad4b2

	PMStandardizationScaler >> scale [
	^ self variance sqrt
	]

To calculate sqrt for every element, compare to 0, and return sqrt or 1.

Other way is to add a handleZeros method, but I do not think we should go there unless we add more types of scaling.

Do you want me to create PR using first suggested fix?

[SergeStinckwich]

Maybe later will have multiple type of scalin, but at the moment, we have only one.
So keep it simple and use first method. Maybe add a test when scale is 0 that it returns 1.