BUG: scatter plot with categorical data raises KeyError

[jorisvandenbossche]

df = pd.DataFrame({'x':[1,2,3,4], 'y':pd.Categorical(['a', 'b', 'a', 'c'])})
df.plot(x='x', y='y', kind='scatter')

raises KeyError: 'y', while the column certainly exists, which can be very confusing.
Without the scatter (just df.plot(x='x', y='y')), it raises the more informative TypeError: Empty 'DataFrame': no numeric data to plot

[stangirala]

@jorisvandenbossche Seems like a simple fix that would check the types of x and y at https://github.com/pandas-dev/pandas/blob/master/pandas/plotting/_core.py#L831?

[TomAugspurger]

@stangirala it may have to be earlier, somewhere around https://github.com/pandas-dev/pandas/blob/master/pandas/plotting/_core.py#L335

The issue is we drop non-numeric columns fairly early on, and by the type you get to https://github.com/pandas-dev/pandas/blob/master/pandas/plotting/_core.py#L831, the data passed to that method already doesn't have the categorical column.

The cleanest way might be to modify MPLPlot._compute_plot_data to check if self.x and self.y are not in numeric_data before setting self.data = numeric_data.

[stangirala]

@TomAugspurger I see. But it looks like PlanePlot has checks on x and y and not MPLPlot, https://github.com/pandas-dev/pandas/blob/master/pandas/plotting/_core.py#L770. The error in the code sample above passes the check in MPLPlot at https://github.com/pandas-dev/pandas/blob/master/pandas/plotting/_core.py#L335 because x is numeric. So the check would go in PlanePlot right?

[jorisvandenbossche]

@stangirala Another interesting area to contribute, however a much bigger issue, is to actually better support categorical data (as the above example could work)

[stangirala]

@jorisvandenbossche Maybe you can open a new issue for it? :D It seems like we would have a flag for most plots that would wholesale convert categorical data to a label-to-integer mapping?

But I don't think it would make much sense for a scatter plot that requires an inherent ordering. But having equivalent categorical plots is a good idea, for example parallel_sets, #12341

[jorisvandenbossche]

There are already some issues about it, eg #12341

But I don't think it would make much sense for a scatter plot that requires an inherent ordering.

Do you mean you think it won't make sense for a scatter plot to support categorical data? I think it could make sense, where you basically use the underlying codes of the categorical as the values to plot

[stangirala]

Oh I meant would someone want a scatter plot for categorical data when most of the times the categories don't have an ordering? I mean in such case won't someone want to use a box plot for example assuming y data is numeric and x is categorical?

[TomAugspurger]

supporting categorical in scatter would be nice when you have a single observation per category. More like a dot plot.

[stangirala]

@TomAugspurger I see, a dot plot would make sense. I don't see an issue for this, do I open one?

[TomAugspurger]

We can use #9069 I think.

…

On Wed, May 3, 2017 at 5:02 PM, Telt ***@***.***> wrote: @TomAugspurger <https://github.com/TomAugspurger> I see, a dot plot would make sense. I don't see an issue for this, do I open one? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#16199 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIkbxYqXIAYLgvVh-FqcdrW-KJNaGks5r2PmTgaJpZM4NOVTt> .