Clarify relationship to ColumnTransformer

[micahjsmith]

Scikit-learn v0.20.0 was just released and includes the ColumnTransformer, which is much related to the DataFrameMapper. It might be helpful for people who come to the sklearn-pandas project to see a note in the README explaining what differences there are between the approaches and when one should be preferred over the other.

[devforfu]

@micahjsmith Yes, agree, it seems like ColumnTransformer makes a lot of work in a similar way. Definitely worth to mention in the README file. Would you like to make a PR with documentation describing the similarities and differences between these two things?

[micahjsmith]

Sure, I can give it a shot

[ganesh-krishnan]

Just checking on this issue. Looks like sklearn now has ColumnTransformer. As such, I'm not sure if there are any additional benefits to using sklearn-pandas. Would someone mind clarifying?

[micahjsmith]

I never followed up above...

But maybe we can start collecting differences on this thread. ColumnTransformer is close to feature parity, and APIs I presume may change. They are quite similar overall with some minor differences.

API differences

functionality	DataFrameMapper	ColumnTransformer
drop unmapped cols	default = False	remainder = 'drop'
drop specific cols	drop_cols = ['A', 'B']	transformer = 'drop'
passthrough unmapped cols	default = None	remainder = 'passthrough'
passthrough specific cols	transformer = None	transformer = 'passthrough'
output dataframe	df_out = True	n/a
apply prefix and suffix	prefix and suffix options	n/a
apply default transformer	default = SomeTransformer()	n/a
global prefix and suffix	prefix and suffix kwargs	n/a
feature naming	user-specified or automatic	user-specified or use make_column_transformer
column selection	str or List[str]	str, array-like of str, int, array-like of int, array-like of bool, slice or callable
treatment of sparse data	only if sparse=True and has sparse output	by default, configurable by sparse_threshold
supervised transformations	yes	yes

Other functionality

• gen_features

Does this look about right? Are we missing anything?

[micahjsmith]

@ganesh-krishnan does the above look about right?

[ganesh-krishnan]

Looks good per my understanding.

I'm not an expert in both by any means. Was struggling on which one to choose. A table like this should be very valuable to folks trying to make a decision.

[arora123]

Few basic differences in DataFrameMapper() and ColumnTransformer(): https://github.com/arora123/Python-for-Data-Science/blob/master/DataFrameMapper_Vs_Column_Transformer.ipynb