Jitter plot

[yoavram]

It would be great if we can have either a jitter command or a jitter flag in scatter.
I'm refering to something like ggplot2's geom_jitter.

A basic solution:

def rand_jitter(arr):
    stdev = arr.max()/100.
    return arr + np.random.randn(len(arr)) * stdev

def jitter(x, y, s=20, c='b', marker='o', cmap=None, norm=None, vmin=None, vmax=None, alpha=None, linewidths=None, verts=None, hold=None, **kwargs):
    return scatter(rand_jitter(x), rand_jitter(y), s=20, c='b', marker='o', cmap=None, norm=None, vmin=None, vmax=None, alpha=None, linewidths=None, verts=None, hold=None, **kwargs)

Of course the definition of stdev needs to be better, this will only work for some cases.

The idea is from this SoF post.

[tacaswell]

I am very against adding a jitter function, I see it as crossing the line to unethical data mangling/falsification (but I know opinions differ). If I saw a colleague doing this I would call them out in very strong terms.

If people really want to do this, than they should do their data-modification explicitly, outside of mpl, so we never get users who are surprised that the library changed their data (in a random way!) before plotting it.

Leaving this issue open just to make I am not way out in left field on this.

[WeatherGod]

I will say that there are some very limited situations where a "jitter"
is a good thing (can help to avoid over-fitting models on low-precision
data and such), I absolutely do agree that such things need to be done
explicitly prior to plotting. One of matplotlib's philosophies is to keep
it simple by just simply plotting the data as given. Failure to separate
the responsibility of preping the data and plotting the data results in
abominations that is the histogram functions.

[tacaswell]

However, in those cases I assume that a) you have a decent model of what the noise should be, b) you run the jitter + fit routine a bunch of times to make sure you get consistent results, and c) are very forth coming about exactly what you did and why you did it.

Plotting the jittered data gives a false sense of precision.

[WeatherGod]

Absolutely agree with those points. However, if you are going to use
jittered data, then I suggest that it helps in explaining that the data was
jittered by plotting the un-jittered data on top of the jittered data, that
way you are completely transparent about it.

On Wed, Jan 22, 2014 at 9:57 AM, Thomas A Caswell
notifications@github.comwrote:

However, in those cases I assume that a) you have a decent model of what
the noise should be, b) you run the jitter + fit routine a bunch of
times to make sure you get consistent results, and c) are very forth coming
about exactly what you did and why you did it.

Plotting the jittered data gives a false sense of precision.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/2750#issuecomment-33030079
.

[efiring]

I agree with the arguments against including a jitter in mpl. The primary legitimate use would appear to be to provide a way of visualizing data sets with duplicate values. Maybe a better way would be to process the data such that the number of duplicates of a given x,y pair would be used to set a color or a symbol size in scatter. In any case, the processing, whether a duplicate count or a jitter addition, can be left as the responsibility of the user, not mpl.

[yoavram]

Thanks for the informative discussion! I didn't think about it that way and I agree with your comments.