[Data/preprocessors] Allow preprocessors to be append operations

[martinbomio]

Description

Today, all preprocessors (or at least most) are expressed as replace operations, meaning that if a preprocessor is applied for column A for a Dataset, it will replace the values of column A with the transformed value.

Allowing the preprocessors to be append operations, creating new columns instead of replacing the input one, allows for more complex preprocessing chains were one can use a column as an input to multiple output columns.

Use case

No response

[richardliaw]

Makes sense; what's the API proposal?

Preprocessor(inplace=True/False)?

[martinbomio]

Regarding API changes, there's a couple of options I can think of:

1. Add a map from original column to new column on preprocessors. Example

   scaler = StandardScaler(columns=["A", "B"], output_column_mapping={"A": "A_scaled", "B": "B_scaled"})
   

   the default behavior can be that if no mapping is provided the transformation is done in place.

2. Add an inplace param to the preprocessors. There will also be the need to add a column_suffix. Example:

   scaler = StandardScaler(columns=["A", "B"], inplace=False, suffix="_scaled")
   

   inplace can default to True for backwards compatibility, suffix can have a sane default per preprocessor. It is a but less
   flexible for the user to fully define the output column name per transformed column.

[richardliaw]

yeah I think (2) I am more comfortable with.

[jcotant1]

Hey @martinbomio @richardliaw where'd this one end up?

[martinbomio]

@jcotant1 I think we informally agreed with @richardliaw to take approach (2) described here

[jcotant1]

Thanks, @martinbomio cc: @gvspraveen