Documentation: Incorrect description of statistic computed by MaxAbsScaler.partial_fit #34900

Introduce yourself

Hi, I’m Pradip, a Computer Engineering student and I’m currently learning and working with Python and machine learning. I use scikit-learn for learning and implementing machine learning algorithms and preprocessing techniques.

While going through the MaxAbsScaler documentation and source code, I noticed a discrepancy in the partial_fit() docstring. I’m opening this issue to clarify the documentation and make it consistent with the actual implementation.

Describe the issue linked to the documentation

The X parameter description in MaxAbsScaler.partial_fit() currently states that the data is used to compute the "per-feature minimum and maximum".

However, MaxAbsScaler computes the maximum absolute value for each feature and stores it in max_abs_.

For example:

X = [[-2, 5],
     [-3, 1]]

For the first feature:

minimum = -3
maximum = -2
maximum absolute value = 3

MaxAbsScaler uses 3 as the scaling value. It does not use the minimum and maximum as separate statistics.

The implementation reflects this:

max_abs = xpx.nanmax(xp.abs(X), axis=0, xp=xp)

Therefore, the current wording may confuse users into thinking that MaxAbsScaler computes and uses both the minimum and maximum values independently.

This also appears inconsistent with the class-level documentation, which describes the scaler in terms of the maximum absolute value.

I believe this is a documentation-only issue and does not require any change to the underlying algorithm.

Suggest a potential alternative/fix

I suggest updating the X parameter description in MaxAbsScaler.partial_fit() from:

The data used to compute the per-feature minimum and maximum
used for later scaling along the features axis.

to:

The data used to compute the per-feature maximum absolute value
used for later scaling along the features axis.

This would make the partial_fit() documentation accurately reflect the statistic computed by the implementation and remain consistent with the class-level documentation.