Description

Right now, sklearn.preprocessing.PowerTransformer implements the Box-Cox transform (#10210), which only works on positive, nonzero data. The Yeo-Johnson transform is an alternative power transformation that supports negative and zero values. Yeo and Johnson also proposed a more robust technique for selecting the transformation parameter (paper linked below), which may be a good alternative to maximum likelihood.

Original paper
Parameter estimation

Comments

@jwilber and I plan to tackle this implementation together.