Update the return dtype of boolean transformer

[katxiao]

In the BooleanTransformer, we always cast the reverse transformed data to type object, because if a boolean series contains NaNs, it will be type object. However, if the input is a boolean series without NaNs (therefore type bool), the output type will be inconsistent.

We could instead try to first cast the output type to bool, and if there is a ValueError due to the presence of NaNs, we could cast it to object. The alternative is to first check if the series has any null values, but we have previously seen that checking for nulls is not very performant.

[codecov-commenter]

Codecov Report

Merging #247 (e07e40d) into master (54e8a0c) will increase coverage by 0.03%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #247      +/-   ##
==========================================
+ Coverage   94.92%   94.95%   +0.03%     
==========================================
  Files           9        9              
  Lines         571      575       +4     
==========================================
+ Hits          542      546       +4     
  Misses         29       29              

Impacted Files	Coverage Δ
rdt/transformers/boolean.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 54e8a0c...e07e40d. Read the comment docs.

[csala]

Good catch @katxiao
This may create conflicts with the changes coming in #228 , so I would wait for that to be merged first and then see if we need to add this on top of it.

On the other hand, I think that the implemented approach is fine but it may also produce inconsistent behaviors: One may have fitted the transformer with nulls (so, passing object as input) but later on be reverse transforming data that does not contain nulls. In this case, I would rather expect the output to be object, like what was seen in fit.

Which actually brings us to a new question: Rather than taking care about this here, at the Transformer implementation level, maybe we should update the new BaseTransformer implementation to capture the fit input dtypes and simply cast back to them at the end of reverse_transform. By doing this, we make sure that any transformer always produces the same dtypes that it saw during fit, which are the ones that are truly relevant.

What do you think about it? CCing @amontanez24 @fealho too

[katxiao]

Yeah, I think ideally the dtype should be set to whatever the input data's dtype is. We can either merge this first, or just wait until we make the larger BaseTransformer change (either is fine with me).