BUG: Allow IntervalIndex to be constructed from categorical data with appropriate dtype

[jschendel]

• closes IntervalIndex does not accept CategoricalIndex (of interval dtype) #21243
• closes IntervalIndex.from_arrays/from_breaks does not accept categorical data with valid dtype #21253
• tests added / passed
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

Added this to 0.23.1 since it's a regression and the fix is a minor change outside the IntervalIndex class. Not opposed to pushing to 0.24.0 if backporting this could be problematic.

[jschendel]

This is being skipped due to the following behavior:

In [2]: cat = pd.Categorical([pd.Interval(0, 1), pd.Interval(1, 2), pd.Interval(0, 1)])

In [3]: pd.Index(cat, dtype='interval')
Out[3]: CategoricalIndex([(0, 1], (1, 2], (0, 1]], categories=[(0, 1], (1, 2]], ordered=False, dtype='category')

This happens because the Index code is structured so that categorical takes precedence over interval:

pandas/pandas/core/indexes/base.py

Lines 262 to 273 in c85ab08

	# categorical
	if is_categorical_dtype(data) or is_categorical_dtype(dtype):
	from .category import CategoricalIndex
	return CategoricalIndex(data, dtype=dtype, copy=copy, name=name,
	**kwargs)
	# interval
	if is_interval_dtype(data) or is_interval_dtype(dtype):
	from .interval import IntervalIndex
	closed = kwargs.get('closed', None)
	return IntervalIndex(data, dtype=dtype, name=name, copy=copy,
	closed=closed)

The code above could be restructured so that the dtype argument, if present, takes precedence over the type of data. Seems like that would be more sensible than the current approach for this corner case, but on the fence about it.

[jreback]

I c. ok can open an issue about this, but yes I would agree should infer with a passed dtype first before switching on the type of the data.

[codecov]

Codecov Report

Merging #21254 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master   #21254      +/-   ##
==========================================
+ Coverage   91.84%   91.84%   +<.01%     
==========================================
  Files         153      153              
  Lines       49538    49540       +2     
==========================================
+ Hits        45499    45501       +2     
  Misses       4039     4039

Flag	Coverage Δ
#multiple	90.24% <100%> (ø)	⬆️
#single	41.87% <50%> (ø)	⬆️

Impacted Files	Coverage Δ
pandas/core/indexes/interval.py	93.16% <100%> (+0.02%)	⬆️

---

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 c85ab08...770fe10. Read the comment docs.

[jreback]

can yoou replicate the test in #21243 in test_tile, otherwise lgtm.

[jreback]

I c. ok can open an issue about this, but yes I would agree should infer with a passed dtype first before switching on the type of the data.

[jreback]

thanks @jschendel