Add YeoJohnson transformer

[tlienart]

Here: https://github.com/tk3369/YeoJohnsonTrans.jl with creds to Tom Kwong

it's in a similar vein to BoxCox

[azev77]

@tlienart this is great!
Here are the pre-processing options I use in Caret:

"BoxCox", "YeoJohnson", "expoTrans", "center", "scale", "range", "nzv",
"knnImpute",  "bagImpute",  "medianImpute",  "pca",  "ica", "spatialSign"

dp= preProcess(d, method = c("center", "scale", "YeoJohnson", "nzv"))

Recently, I've found bestNormalize helpful.

[tlienart]

Cool! could you explain what

• nzv (non zero value?)
• bagImpute
• spatialSign

do ?

(also could you link to relevant doc from Caret?)

Finally bestNormalize looks great, if someone felt up to coding something like that in Julia it'd be great to interface with it

[azev77]

https://github.com/topepo/caret/blob/master/pkg/caret/R/preProcess.R

https://github.com/topepo/caret/blob/master/pkg/caret/R/BoxCoxTrans.R

https://github.com/topepo/caret/blob/master/pkg/caret/R/expoTrans.R

https://github.com/topepo/caret/blob/master/pkg/caret/R/nearZeroVar.R

https://github.com/topepo/caret/blob/master/pkg/caret/R/spatialSign.R

[azev77]

@tlienart btw
Several of the preprocessing options above are for imputing missing data.

MLJ comes w Ames housing data, they already imputed missing values & cleaned the data. Do you know which methods they used?

I think The most popular is:
https://github.com/stefvanbuuren/mice
(I wish it was faster...)

Julia has:
https://github.com/invenia/Impute.jl

You may have seen: https://discourse.julialang.org/t/how-to-do-multiple-imputation-on-julia/17713/14

[tlienart]

So we have imputation mechanisms here (https://github.com/alan-turing-institute/MLJModels.jl/blob/master/src/builtins/Transformers.jl) which currently allows you to do whatever you want that is column based (e.g.: median/mean imputation).

More intricate imputation becomes a full blown transformer, and we could just interface with any package that would provide such a thing. Interfacing with Impute.jl would be nice, they don't offer much more than what we already do but do offer carry forward for instance (which we could easily also implement here). Assuming impute.jl will grow over time it's probably still a good idea to interface with it.

(maybe open another issue for discussion of imputers?)

[azev77]

Impute.jl is discussing implementing something like mice
invenia/Impute.jl#3