Histogram gap artifacts

[fonnesbeck]

I'm hoping there is a way to improve the default behavior of hist, which often generates gap artifacts between classes in histograms. For example, even using clean integer data that has natural breaks:

lsl_dr.degree_hl.dropna().value_counts()

6    10446
3     3127
4     3041
5     2705
2     1416
0      549
1      170
dtype: int64

I get a result like the following:

This plot is misleading because it suggests there are intervening classes with zero counts in between the bars. I assume this is a solved problem, since it does not occur in other packages (e.g. ggplot).

[fonnesbeck]

I'm going to try my hand at a pull request for this. Stay tuned.

[tacaswell]

Pretty sure the issue is that the default number of bins in 10. We don't currently try to do anything intelligent on guessing at 'good' values for bins.

Also see #4316

[efiring]

On 2015/04/15 3:54 AM, Chris Fonnesbeck wrote:

I'm going to try my hand at a pull request for this. Stay tuned.

Chris, check the numpy mailing list. The pull request for this should go
to numpy, not matplotlib. There is a discussion going on now about
this, on the numpy list. Moving that along, to ensure the functionality
gets into numpy.histogram soon, would be good. (It might be in danger
of being delayed or derailed by overly complex proposals.)

If auto bin selection is added to numpy.histogram, we will support it in
matplotlib. Depending on how complicated it is we could consider
including a backport to handle the case of people using up-to-date
matplotlib with older numpy.

[tacaswell]

closing as not-a-matplotlib problem.

A version of numpy with 'auto' as a valid bin argument is now out (but I don't think it works well with categorical axes yet).