ENH: Sparse int64 and bool dtype support enhancement

[sinhrks]

• closes #667, closes #8292, closes #13001, closes #8276, closes #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

[sinhrks]

rebased and added the doc, now ready for review.

[jreback]

make this a separe sub-section (as first one is too long?)

[jreback]

could add a ref to #667 here as well

[jreback]

lgtm. just some doc corrections. ping on green.

[jreback]

say dtype='category' for parsing Categorical data.

[jreback]

I think explicty put the supported dtypes in here to have a nice error message, for now to avoid unwanted conversions (e.g. this won't raise on M8, and add a test for non-supported dtypes.

[jreback]

prob need to have this handled for .astype as well. you already have something like this getting the fill value, maybe can use that to validate. (e.g. use the if test, then raise if its not there).

[jreback]

actually maybe should put this in sparse.array or have a sparse.common module that you can import to do centrailzed sparse things .

[jreback]

elif is_floating_dtype(array_or_dtype):
    return np.nan

raise ValueError("unsupported dtype {...} for Sparse")

[jreback]

add a clause for object, raise otherwise

[sinhrks]

@jreback Thx for review. One point is whether we should prohibit dtypes other than "relatively-well" supported ones (currently float64, after the PR int64 and bool also). Maybe my title was misleading, thus change the title.

There are few issues which uses sparse data with object dtype like #11633, #13917. So I feel we should keep the current behavior for other dtype rather than raise (may better to show a warning).

CC: @sstanovnik

[jreback]

yeah object dtypes are only partially supported ATM. I think can leave that ok, though possibly we could warn, @jorisvandenbossche ?

[sstanovnik]

Let me add my 2¢, since you went as far as CC-ing me. I can't make a very informed opinion, since I don't know enough about pandas' internals, and I obviously have an interest (biolab/orange3#1347) for supporting arbitrary types.

My thoughts are that you should be able to throw the same kind of data in a dense or a sparse DataFrame so that they are equivalent. An example off the top of my head is a SparseDataFrame with a recommendation dataset with rows as movies and columns as users, and additional metadata (string) columns about each movie. I don't know if this is possible, but judging from my time with BlockManager, you could maybe use dense string columns mixed in-between an otherwise sparse structure, if supporting sparse string storage is too hard.

As I said, I may be completely off-target here, just some thoughts :)

[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 pydata#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

[jreback]

need a sub-section ref

[jreback]

structure -> structures

[jreback]

have -> were
even if all inputs are -> even if all inputs were int or bool dtype.

Unclear what the Also, you must specify sentence means....can you reword?

[jreback]

I am not sure about this. I think if a float fill_value should force float (though if the data infers as integer, maybe raise/warn)? having the determination as no gaps is too data specific. I know this forces an integer array to then have a specified fill value (as it can't then be np.nan), but I think that's ok.

[jreback]

my remaining question was this one

[jorisvandenbossche]

I agree that it seems logical that the fill_value matches the dtype. So in case of specifically specifying the fill_value, I would take that into account for the actual dtype inference.

Given that nan is no longer the default fill_value, I don't think it is a problem that specifying fill_value=np.nan changes your integer data into a float sparse array.

[jorisvandenbossche]

@sinhrks The docstring of SparseArray still says that the default fill_value is NaN, which is no longer true I think (it changed to None, to depend on the data type I suppose)

[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 ?

[jorisvandenbossche]

I don't think the % _shared_doc_kwargs are used in this case?

[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.