[WIP] ENH: Groupby mode

[lithomas1]

• closes Groupby.mode() - feature request #19254
• tests added / passed
• Ensure all linting tests pass, see here for how to run them
• whatsnew entry

Proof of concept for #19254

Is relatively fast compared to @rhshadrach method

In [4]: %timeit df.groupby("a").mode()
2.52 ms ± 415 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [5]: %timeit df.groupby("a").mode()
1.99 ms ± 60.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [6]: %timeit gb_mode(df, ['a'], 'b')
3.9 ms ± 310 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [7]: %timeit gb_mode(df, ['a'], 'b')
3.8 ms ± 270 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [8]: %timeit gb_mode(df, ['a'], 'b')
3.89 ms ± 433 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

API for what to do when multimodal needs discussion though.

[rhshadrach]

Thanks @lithomas1! Can you include the frame used for the time test? If the values don't have high cardinality (> 500 distinct values?), can you also test with one that does? My expectation is that the cython implementation performs much better the larger the distinct values are.

[jreback]

is this not just value counts(sort=True, dropna=True)[0]? why are we adding a whole bunch of cython code

[lithomas1]

Well value_counts is about twice as slow as my current implementation(takes same time for 1 column as my implementation does for 2), and all the other methods have corresponding cython code.

[rhshadrach]

What is the code that can do groupby.mode using value_counts?

[rhshadrach]

Ah - whoops, I was trying to use DataFrameGroupBy instead of SeriesGroupBy. Ignore the above.

[jreback]

Well value_counts is about twice as slow as my current implementation(takes same time for 1 column as my implementation does for 2), and all the other methods have corresponding cython code.

can you show some timings. I am really hesistant to add this much code for a niche case.

[jbrockmendel]

using cython fused types should make cython figure this out for us (also avoids need for tempita)

[lithomas1]

I have to use tempita unfortunately because of the hashtables have different functions names based on dtype.

[WillAyd]

What is the expected result for column B when doing this?

[WillAyd]

To @jbrockmendel point a lot of this can be simplified if using fused types instead of a template; we've started to move to that in most other places in Cython

[WillAyd]

The non-numeric code paths prevent the Gil from being released. At least until Cython 3 gets released we usually do something like this:

pandas/pandas/_libs/groupby.pyx

Line 928 in 859a787

if rank_t is object:

[lithomas1]

Dont have time for this.