Fix SimpleImputer error which occurs when all features are bool type

[christopherbunn]

Fixed issues with all-bool inputs for Imputer and SimpleImputer by internally converting to the category datatype and reconverting to bools at the end.

Resolves #1166

[codecov]

Codecov Report

Merging #1215 (c47f85d) into main (b213240) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@            Coverage Diff            @@
##             main    #1215     +/-   ##
=========================================
+ Coverage   100.0%   100.0%   +0.1%     
=========================================
  Files         223      223             
  Lines       15025    15100     +75     
=========================================
+ Hits        15018    15093     +75     
  Misses          7        7             

Impacted Files	Coverage Δ
...components/transformers/imputers/simple_imputer.py	100.0% <100.0%> (ø)
evalml/tests/component_tests/test_imputer.py	100.0% <100.0%> (ø)
...valml/tests/component_tests/test_simple_imputer.py	100.0% <100.0%> (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 b213240...c47f85d. Read the comment docs.

[jeremyliweishih]

cool looks good!

[dsherry]

@christopherbunn thank you, looks good! I left one question about the behavior of the imputer for boolean inputs, may be a bug there.

[dsherry]

Why were nans here filled with True rather than False? Your example data had one True and two False. I thought the default categorical strategy of most_frequent would fill nans with False.

[dsherry]

@angela97lin heads up these changes will cause conflicts for #1423 (update all components to accept woodwork datatables). I think they should be pretty straightforward to address--when you're resolving the conflicts you could call dt.set_logical_types(...) to do this transformation, or you could keep doing the transformation with pandas. Great that @christopherbunn is adding test coverage for this case, that'll help!

[dsherry]

Ah, #1423 was just merged!

@christopherbunn if you need a hand updating this, please ping @angela97lin or I

[angela97lin]

Oo yup, just saw this. Yes yes, please reach out if you need help fixing the merge conflicts!

[christopherbunn]

@dsherry Thanks for the review! I looked at the test case a bit deeper and part of the reason why that passes is that np.nan actually gets casted to True when the data type is set to bool. As an example, this snippet below

    X_multi = pd.DataFrame({
        "bool with nan": pd.Series([True, np.nan, False, np.nan, False], dtype=bool),
        "bool no nan": pd.Series([False, False, False, False, True], dtype=bool),
    })

is actually equal to

    pd.DataFrame({
        "bool with nan": pd.Series([True, True, False, True, False], dtype=bool),
        "bool no nan": pd.Series([False, False, False, False, True], dtype=bool),
    })

Because of this, I'm not actually sure if its possible to impute a bool-only df like X_multi since it will always be filled with values that are either True or False. X_multi could exist with True, False, and np.nan values but only if the dtype is Categorical from the beginning. In that case, the existing code would already work.

I think that resolving this issue should instead focus on just catching this value error. The current HEAD of this PR will just return all-bool dfs as is without imputing it and I added a new case to test this behavior. Thoughts @angela97lin and @dsherry?

[angela97lin]

@christopherbunn Yup, I'd agree that resolving this issue to just catch the ValueError sounds good to me. It's a limitation or design choice made by pandas, and one I came across as well while implementing Woodwork stuff. Now that we have woodwork integration in place for components, users can use Woodwork structures, which use the new nullable pd.NA, and this would be converted to a category (or pandas "object") column, which you've mentioned already works. (Maybe this is a worthwhile test to add for the imputer specifically?)

[dsherry]

Discussed with @angela97lin @christopherbunn . Recommendations:

• For SimpleImputer, if we get this error, catch the error and return the original data
• Unit tests should cover both pandas as input and woodwork as input

[dsherry]

I left a few suggestions which we should resolve before merging but LGTM!

[dsherry]

This test is great!

I think the first two cases (i.e. above this line) are fully covered by the last case, right? I suggest you delete the first two cases, above this line, and keep the last one below.

[christopherbunn]

I'd argue for keeping all of the test cases. The first test case triggers the type conversion while the second one checks for proper boolean impution. The third one is more of an extension of the second one but it also combines a complete boolean column to make sure that the types are being processed properly.

[dsherry]

Suggest "Fix SimpleImputer error which occurs when all features are bool type"

[freddyaboulton]

@christopherbunn I think this is fantastic! Thanks for making this change. I think your unit test for the Imputer shows that SimpleImputer and Imputer convert dtypes between input and output differently (even though Imputer calls SimpleImputer under the hood).

I don't think that should block merge (maybe that behavior existed before this change) so maybe the best thing is to address @dsherry 's comments (and any others that come up) and file an issue to continue the discussion.

[dsherry]

Unit tests discussed with @christopherbunn

1. All cols are bool-type
   X = pd.DataFrame([True, True, False, True, True], dtype=bool)
2. Imputation strategies work for bool type
   X = pd.DataFrame([True, np.nan, False, np.nan, True], dtype=object)
X = pd.DataFrame([True, np.nan, False, np.nan, False], dtype=object)
3. Multi-type, with at least one bool typed col

    X_multi_expected_arr = pd.DataFrame({
        "bool with nan": pd.Series([True, False, False, False, False], dtype=object),
        "bool no nan": pd.Series([False, False, False, False, True], dtype=bool),
    })

[dsherry]

@christopherbunn should this comment be in fit too?

[dsherry]

@christopherbunn thanks for chasing this down!