Stacked bar graph: Bar segments are sorted in opposite order of legend.

[csaid]

#The legend is sorted 'A, B, C', but the bar segments are sorted 'C, B, A'.
d = {'A': [2,3], 'B': [4,5], 'C': [6,7]}
df = pd.DataFrame(data=d)
df.plot(kind='bar', stacked=True)

Thanks, and here's my full version string:

INSTALLED VERSIONS

Python: 2.7.6.final.0
OS: Darwin
Release: 13.0.0
Processor: i386
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8

pandas: 0.13.0
Cython: Not installed
Numpy: 1.8.0
Scipy: 0.13.1
statsmodels: Not installed
patsy: Not installed
scikits.timeseries: Not installed
dateutil: 2.2
pytz: 2013.9
bottleneck: Not installed
PyTables: Not Installed
numexpr: Not Installed
matplotlib: 1.3.1
openpyxl: Not installed
xlrd: Not installed
xlwt: Not installed
xlsxwriter: Not installed
sqlalchemy: Not installed
lxml: Not installed
bs4: Not installed
html5lib: Not installed
bigquery: Not installed
apiclient: Not installed

[bjornarneson]

This seems like the expected behavior. A stacked column or bar chart should plot the first variable nearest to the intersection of the X & Y axes, right?

[csaid]

From top to bottom the sequence in the legend is ABC, but the sequence in the bar segments is CBA. I don't have a strong opinion about which sequence is correct, but I definitely think there should be consistency between the sequence in the bars and the sequence in the legend.

[bjornarneson]

I appreciate what you are saying, but if you think about the order of the segments & legend as 'first to last' rather than 'top to bottom', the current behavior makes sense. The first value is always the one nearest to the x-axis.

Imagine if the values of a, b, and c were all negative. 'Top to bottom' takes on a different meaning on the other side of the x-axis...or if the dataset was rendered as a horizontal bar chart.

[csaid]

Thanks for the comment, although I totally disagree :)

Stacked bar graphs are hardly ever negative. In the typical case they are positive, and it makes it a lot easier for people to match segments to labels if they follow the same order in the 'top to bottom' sense (i.e the visually intuitive sense), not the 'first to last' sense.

I'm new to issue reporting on github and so I don't know the etiquette for these things, but I'm curious if we can get someone else's opinion on this.

[ghost]

Let's try a constructional proof:

• The quantities to be displayed have a well-defined ordering.
• Reading from top to bottom is quite popular.
• We must stack things in bottom-up order. I learned this the hard way
  playing "hanoi towers" as a small child. It was tough going there for a while.
• The reading order of items in the legend should conform to the ordering on the quantities.
• The stacking order should conform to the ordering on the quantities.
• Since the legend is ordered top to bottom and the bars are stacked bottom to top,
  they express the same ordering on the quantities in the plot as shown, and the plot
  is therefore consistent. ∎

That's not convincing in the least, but it shows the behavior makes sense for some
definition of sense.

Show me a major piece of related software that does this differently: ggplot2,
excel, tableau or one of the big stats packages and we'll seriously consider changing it.

Demonstrate that there's In fact a consensus among most of those and this ain't it, and
we will change it.

The etiquette on GH issues is to argue your point well while remaining polite. Amuse if you're able.

[csaid]

Hi all,
I probably shouldn't have used 'A, B, C' as my example because those already have a well defined ordering. Where this really matters is when you have categorical variables without a well-defined ordering. Here is default behavior in pandas. I think this is hard to read.

Here is the default behavior in Excel, which I think is much easier to read. I also like how Excel preserves the order of the categories rather than sorting alphabetically.

I personally think the Excel-style visually-matching ordering should be the default, but if that's not possible do you think there could be a simple "reverse labels" option in the plot method, or some other simple way to accomplish this? Thanks.

P.S, I don't have Tableau, but I did a google image search, and it looks like the segments and labels are sorted in the same way visually, as in Excel.
https://www.google.com/search?q=tableau+stacked+bar+graph&tbm=isch

[ghost]

I've opened a PR for 0.14.0, we'll see how it goes. the plotting part of pandas is really fragile.
Also, I suggest you make use of pd.options.display.mpl_style='default', that color scheme
is really tragic.

[csaid]

Thanks y-p, much appreciated.

[ghost]

I just merged #6118.