ENH: Sparse int64 and bool dtype support enhancement

[sinhrks]

• closes Support dtypes other than float in sparse data structures #667, closes BUG: int sparse series is converted to float if sliced #8292, closes _combine_const() in pandas.sparse.frame does not have uniform method signature with pandas.core.frame #13001, closes SparseDataFrame.isnull raises an error #8276, closes BUG: SparseSeries/DataFrame non-float dtypes repr #13110
  • tests added / passed
  • passes git diff upstream/master | flake8 --diff
  • whatsnew entry

Currently, sparse doesn't support int64 and bool dtypes actually. When int or bool values are passed, it is coerced to float64 if dtypekw is not explicitly specified.

on current master

pd.SparseArray([1, 2, 0, 0 ])
# [1.0, 2.0, 0.0, 0.0]
# Fill: nan
# IntIndex
# Indices: array([0, 1, 2, 3], dtype=int32)

pd.SparseArray([True, False, True])
# [1.0, 0.0, 1.0]
# Fill: nan
# IntIndex
# Indices: array([0, 1, 2], dtype=int32)

after this PR

The created data should have the dtype of passed values (as the same as normal Series).

pd.SparseArray([1, 2, 0, 0 ])
# [1, 2, 0, 0]
# Fill: 0
# IntIndex
# Indices: array([0, 1], dtype=int32)

pd.SparseArray([True, False, True])
# [True, False, True]
# Fill: False
# IntIndex
# Indices: array([0, 2], dtype=int32)

Also, fill_value is automatically specified according to the following rules (because np.nan cannot appear in int or bool dtype):

Basic rule: sparse dtype must not be changed when it is converted to dense.

• If sparse_index is specified and data has a hole (missing values):
  • fill_value is np.nan
  • dtype is float64 or object (which can store both data and fill_value)
• If sparse_index is None (all values are provided via data, no missing values)
  • if fill_value is not explicitly passed, following default will be used depending on its dtype.
    • float: np.nan
    • int: 0
    • bool: False

[codecov-io]

Current coverage is 85.27% (diff: 98.63%)

Merging #13849 into master will increase coverage by <.01%

@@             master     #13849   diff @@
==========================================
  Files           139        139          
  Lines         50511      50523    +12   
  Methods           0          0          
  Messages          0          0          
  Branches          0          0          
==========================================
+ Hits          43071      43083    +12   
  Misses         7440       7440          
  Partials          0          0          

Powered by Codecov. Last update 10bf721...341585a

[jreback]

@sinhrks getting tons of warnings compiling on windows....all the same

pandas\src\sparse.c(63861) : warning C4244: '=' : conversion from 'Py_ssize_t' to '__pyx_t_5numpy_int32_t', possible loss of data
pandas\src\sparse.c(63870) : warning C4244: '=' : conversion from 'Py_ssize_t' to '__pyx_t_5numpy_int32_t', possible loss of data
pandas\src\sparse.c(66180) : warning C4244: '=' : conversion from 'Py_ssize_t' to '__pyx_t_5numpy_int32_t', possible loss of data
pandas\src\sparse.c(66189) : warning C4244: '=' : conversion from 'Py_ssize_t' to '__pyx_t_5numpy_int32_t', possible loss of data
pandas\src\sparse.c(68499) : warning C4244: '=' : conversion from 'Py_ssize_t' to '__pyx_t_5numpy_int32_t', possible loss of data

[jorisvandenbossche]

Sorry, not familiar with sparse. But: using object dtype, does it work enough to use it for certain cases? If yes, I would not remove it.

[jorisvandenbossche]

@sinhrks Does this also close #13110?

[sinhrks]

I think object dtype can be used in some cases, but not fully sure as it is not tested well. Not remove ATM and add more tests to clarify (on another PR).

#13110 should be closed. Added whatsnew.

[jreback]

would leave out other

[jorisvandenbossche]

Disclaimer: I never used sparse or am familiar with the implementation (so my excuses if it is a stupid or naive question), but I quickly looked at the PR and have the following question.

Previously, for integer and boolean serieses, the 0 or False values were regarded as actual values, not an indication of 'not a value' in the sparse series. Isn't this a big change? (I don't know how much you could use it before this PR to be a problem)
Next to that, having eg False for boolean arrays as the default fill_value also seems a bit strange to me. I would expect that somebody who wants a boolean sparse array, would want to be able to have both True and False values as actual values? (eg something like [True, -, -, False, -, -, True])?
Of course this is currently because boolean serieses cannot have anything else as True or False.

[jorisvandenbossche]

OK, so probably my question should be categorized in the naive category :-)
I see that this is the same as what scipy.sparse does, so seems like a sensible default then.

[jorisvandenbossche]

Can you add a note here somewhere that for int and bool this was only added from 0.19 ?

[jreback]

joris your example already works you can have any values u want as actual values (both True and False); the fill value is for the missing value indicator when I need to densify (it's the default)

so this is not a conceptual change at all just a change to keep dtype consistency

[jorisvandenbossche]

@jreback I was looking at the to_sparse examples. So the fill_value is also used to convert from dense to sparse. So the output what you see there (eg in case of pd.Series([1, 0, 0]).to_sparse()) has changed (previously that was a block length of 3, now of 1). But no problem, I understand that the actual behaviour you want has not changed.

[jorisvandenbossche]

@jreback This PR for the rest OK to merge for you, Jeff? (it's closing a lot of issues for 0.19.0 :-))

[jorisvandenbossche]

@sinhrks Can you update the docstrings for SparseDataFrame, SparseSeries and SparseArray? They all still mention the fact that only floats are supported or that nan is the default fill value.

[jorisvandenbossche]

@sinhrks Thanks a lot!

[jorisvandenbossche]

@sinhrks appveyor started failing (some int dtype issues):

======================================================================
FAIL: test_append_zero (pandas.sparse.tests.test_list.TestSparseList)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\Python27_64\envs\pandas\lib\site-packages\pandas\sparse\tests\test_list.py", line 64, in test_append_zero
    tm.assert_sp_array_equal(sparr, SparseArray(arr, fill_value=0))
  File "C:\Python27_64\envs\pandas\lib\site-packages\pandas\util\testing.py", line 1392, in assert_sp_array_equal
    assert_numpy_array_equal(left.sp_values, right.sp_values)
  File "C:\Python27_64\envs\pandas\lib\site-packages\pandas\util\testing.py", line 1083, in assert_numpy_array_equal
    assert_attr_equal('dtype', left, right, obj=obj)
  File "C:\Python27_64\envs\pandas\lib\site-packages\pandas\util\testing.py", line 878, in assert_attr_equal
    left_attr, right_attr)
  File "C:\Python27_64\envs\pandas\lib\site-packages\pandas\util\testing.py", line 1018, in raise_assert_detail
    raise AssertionError(msg)
AssertionError: numpy array are different
Attribute "dtype" are different
[left]:  int64
[right]: int32

[sinhrks]

@jorisvandenbossche thx for pointing out, will fix.