Allow normalizers to skip NaN values

[27pchrisl]

Hi,

I have a dataset with a lot of missing data, and I wanted to use the HotDeckImputer with the Gower kernel to fill in the blanks. I had preprocessed my dataset so that nulls were converted to NAN or ? depending on the data type.

The problem was then that Gower expects continuous features to have been normalized. I then wanted to use the MinMaxNormalizer to do this on the continuous features in the dataset, but it doesn't handle NAN - essentially every value is normalized to zero.

I updated the MinMaxNormalizer and the MaxAbsoluteScaler to skip NAN values, and compute min/max or abs only the finite values and leave the NAN values where they were in the original dataset.

Being new to ML, I wasn't sure if this was a valid approach for using the normalizers together with the Gower imputer - feedback welcome!

[andrewdalpino]

Targetting ML 3.0 release with this since it can be construed as a backwards compatibility break.