Hi @chanansh !

I integrated a modified version of your FeatureUnion-based pipeline into the current version of DataFrameMapper in https://github.com/paulgb/sklearn-pandas/tree/feature_union_pipe. Let me know what you think.

mapping_to_pipeline doesn't seem to check for column_name uniqueness, which leads to strange behavior if the same column is used twice. Alternatively, it seems this functionality could be added to the current implementation of DataFrameMapper if it was modified to accept step names, similarly to sklearn.pipeline.Pipeline. Unfortunately I don't see an easy way to do that. Maybe check the length of the mapping tuples and if no name is provided, automatically generate one such as "col_trans_0". This shouldn't break backwards compatibility.
@chanansh @dukebody

Good idea! I believe we can generate feature names from the selected columns, like it's done in the PR now, and allow the user to provide a custom name as a third argument to the feature tuple.

For example:

mapper = DataFrameMapper([
    (['height'], StandardScaler()),
    (['height'], None, 'unmodified_height'),
])

Feature names (both custom and auto-generated) can be checked during init to ensure they are unique, and raise an exception otherwise.

Does that sound reasonable?

Yeah, that looks like a good approach. I can work on this, but I'm very busy these days so it might be a few weeks until I have time to start.

Don't worry, I'm already working on this myself. Will upload updated code this evening.

Will this be implemented in the master any time soon?

Hi @havardl . We worked on this like one year ago, but honestly I don't think it adds so much value in terms of features compared to the amount of flexibility lost and the fact that it would break all previous pickled mappers (it's backwards incompatible).

I believe it would be better to try to implement the feature of setting parameters as mentioned in the original post of this PR.