Violin plots?

[olgabot]

Besides boxplots, another way to look at the distribution of data is by violin plots:

http://nbviewer.ipython.org/gist/olgabot/7902901

I'd like to add this to pandas. Is there interest in these plots? It currently depends on my prettyplotlib but that can be removed. My concern is that it depends on scipy.stats.gaussian_kde - is that fine?

[TomAugspurger]

Statsmodels has them. I think the idea is to keep the more statistical things there/maybe move some pandas stats functionality over.

[hayd]

These look fun! I think that dep is fine (similar to elsewhere in codebase: https://github.com/pydata/pandas/blob/0327addef295fef292f2fdf8c95546c4cc039abb/pandas/tools/rplot.py#L528)

Edit: I posted this before seeing @TomAugspurger's comment...

[jtratner]

except that the rplot module isn't very well maintained and is quite slow.

[olgabot]

Ah yes statsmodels. I looked at it but didn't end up using them because you couldn't specify the kernel bandwidth.Theirs is quite nice because you can do the left/right distributions for direct comparisons. Maybe I'll just add a PR there that adds a bandwidth specification to the args.

[jtratner]

If they already have some of this, probably better to move it to there
instead. Statsmodels has pandas as a dep so you can still use pandas stuff.

[ghost]

Note that rplot, in retrospect, was a controversial merge precisely for the reason @TomAugspurger mentioned.
So that's indeed the guideline we've tried (or should have) to stick to in the past.

@jtratner , yep, it was in PR limbo for the longest while and I wasn't aware of the statsmodels issue
which is why I prodded for it's merge. It was, in fairness, a sizeable piece of code to contribute. #2015, statsmodels/statsmodels#692

[olgabot]

Okay great! I'll stick to statsmodels for the violinplots. Should other stuff like the clustergrams/clustered heatmaps also be there?

[ghost]

It's a fuzzy boundary. The plots are beautiful but are far more bespoke then pandas' current meat
and potatoes stance, they do stick out a bit among the plaid and denim.

[ghost]

We might call them "2nd order plots" maybe? :)

[jreback]

maybe its worthwhile to refactor out these secondary plots from both pandas and statsmodels and make a common import? (so in effect available in both, but not code dupe)

[ghost]

Where's the all-singing, all-dancing pydata exploratory data viz library??

[olgabot]

pydataviz? aka every data analyst's ipython notebook and codebase...

[dragoljub]

I'm hoping https://github.com/vispy/vispy will be the beginning of such a plotting library. Plotting 100M points a second with a GPU beats the 25 seconds plus it takes for matplotlib right now. 😄

[ghost]

@olgabot, I see ggplot2 supports these plots as a distinct geom, would be nice to have these
available in http://github.com/yhat/ggplot. I'm rooting for them.

cc @JanSchulz

[ghost]

#783 still needs a champion, that plot should fit right in to pandas.

[jtratner]

@y-p @cpcloud @jreback and others - so what's the deal with this PR, dendogram PR, #5700, and #783? Are those things in scope for pandas? I'm not clear from the discussion where we stand on that...

[olgabot]

Violin plots I'm working on in statsmodels so I'll close this one.

[ghost]

@jtratner, no decrees here - just opinion. I think #783 is a reasonable addition to include in pandas,
while the dendogram and violin plots are great but belong in a more specialized context.
Also, I think it'd be a good idea to pool these efforts towards building up viz capabilities
behind a single project that plays nicely with pandas.
Not necessarily a pandas dependency, mind you, just as long as it's easy to use them together.

Please feel free to disagree loudly.

[ghost]

Noting that these seem to have landed in seaborn, and are part of ggplot and so
should land eventually in the python port if they haven't already.

Relieved I'm not just shooting down functionality. These actually find their way to
a better home... :)