API/DEPR: deprecate SparseSeries.from_coo and accept in constructor

[jreback]

xref #15497

using SparseDataFrame as the model.

[jreback]

cc @kernc

[kernc]

So how would this work? The exact same functionality as currently in .from_coo which assigns a multiindex? Note, series from_coo is currently different from any series sliced from a sparse data frame:

>>> spm = scipy.sparse.coo_matrix(np.eye(5))

>>> pd.SparseSeries.from_coo(spm)
0  0    1.0
1  1    1.0
2  2    1.0
3  3    1.0
4  4    1.0
dtype: float64
BlockIndex
Block locations: array([0], dtype=int32)
Block lengths: array([5], dtype=int32)

>>> pd.SparseSeries.from_coo(spm.tocsr()[0].tocoo())  # first row only
0  0    1.0
dtype: float64
BlockIndex
Block locations: array([0], dtype=int32)
Block lengths: array([1], dtype=int32)

>>> pd.SparseDataFrame(spm).iloc[0]  # first row sliced from df
0    1.0
1    NaN
2    NaN
3    NaN
4    NaN
Name: 0, dtype: float64
BlockIndex
Block locations: array([0], dtype=int32)
Block lengths: array([1], dtype=int32)

[jreback]

Looking at this again, I think we should just fully deprecate this method. You can do the exact same thing with SparseDataFrame and its more natural. Though I can see a MultiIndex with a set of data is pretty much what COO is.

I could see SparseSeries accepting a 1-d sparse structure though (is there a concept in scipy)?

Maybe have a read in the original issue and see that the use was.

[kernc]

a MultiIndex with a set of data is pretty much what COO is

But that's an implementation detail of what is generally an arbitrary n-d structure. (row, col), value triplets could just as well (or better) be represented with a non-sparse Series.

I could see SparseSeries accepting a 1-d sparse structure though (is there a concept in scipy)?

All scipy sparse matrces are ndim == 2. Since spm.mean(0) is a 1x5 matrix, we still might accept such a matrix in the constructor when any dimension has length 1. Will prepare something. Is it too late to get it in v0.20.0?

---

to_coo/from_coo introduced in PR #9076 fixing #8048.
The main use case seems to have been the use of SparseSeries.unstack() to shape the MultiIndex sparse series into a sparse data frame. So indeed, using the frame is more natural.

[jreback]

@kernc yeah the more I look at this the more I think we should just remove these from Series entirely. Is there truly a useful case for having them on Series?

[jreback]

Will prepare something. Is it too late to get it in v0.20.0?

yes; to the extent that something is a true bug fix or simply an enhancement it can go in a point release (e.g. 0.20.1)