Plotly.py: px - string or categorical integer values considered as numeric in strip/box/violin plots

[DrGFreeman]

Issue:

When using integer values as categorical variable in a strip / box / violin plot, the values of the categorical variable are mapped to a continuous numeric axis even if the values are of string or pd.Categorical type.

Example:

We create a dataframe with columns names having an integer value in string format. These could be any category that makes sense to the specific business case (e.g. product code, etc.)

import numpy as np
import pandas as pd
import plotly_express as px

n = 50

df = pd.DataFrame({
    '1': np.random.normal(2, .3, n),
    '2': np.random.lognormal(.5, .2, n),
    '34': np.random.triangular(0, 2, 3, n),
    '123': np.random.uniform(1, 3, n)
})

We unpivot the data using and make a strip plot.

df1 = df.melt()

px.strip(df1, x='variable', y='value')

The categorical variable (values '1', '2', '34' and '123') get mapped to a continuous numeric scale. Here, the variables '1' and '2' blend together and this can get worst if there are orders of magnitude between the different values.

Converting the string values to pd.Categorical type yields the same result as above.

df2 = df.melt()
df2.variable = pd.Categorical(df2.variable)

px.strip(df2, x='variable', y='value')

Workaround:

Adding a character to the values makes them be recognized as categorical which is the expected result (except for the added character in the category names). Unfortunately, adding a blank space does not work either.

df3 = df.copy()
df3.columns = [f"c{c}" for c in df3.columns]

px.strip(df3.melt(), x='variable', y='value')

Considering that numeric categorical values are legitimate in many contexts, it should be possible to use numbers as categories if they are represented by a string or categorical data type, as is the case with the color parameter (https://github.com/plotly/plotly_express/issues/140):

px.strip(df.melt(), y='value', color='variable')

Thanks!

Package              Version  
-------------------- --------- 
plotly               4.1.1    
plotly-express       0.4.1    

@emmanuelle, this is one of the two issues we discussed at the PyData meetup.

[emmanuelle]

Hey @DrGFreeman sure this is a valid concern. You can force the axis to be categorical by creating the plotly figure using the px function and then do

fig.update_layout(xaxis_type='category')

which will force your axis to be categorical. Then of course there is the question whether we should impose this at the plotly.express level...

[DrGFreeman]

Thanks for the quick response @emmanuelle. I will use this tip for sure.

Then of course there is the question whether we should impose this at the plotly.express level...

Of course, that is up to the core developers to decide. From the perspective of API consistency, if integers passed as strings to the color parameter yield a categorical color scale (as opposed to a continuous one), I would intuitively expect to get similar results on an axis.

[cvrnogueira]

@emmanuelle thanks for your solution, but when we are in a plot using facet_col this just work for the last plot, not for all of them. Do you have any solution? Thanks

[nicolaskruchten]

@cvrnogueira If you use fig.update_xaxes(type='category') it will apply to all facets.

[nicolaskruchten]

Of course, that is up to the core developers to decide. From the perspective of API consistency, if integers passed as strings to the color parameter yield a categorical color scale (as opposed to a continuous one), I would intuitively expect to get similar results on an axis.

I agree, but for backwards-compatibility reasons throughout the API, we can't really change this at the moment: we've always accepted stringified numbers as numbers on positional axes like x and y