Can plotnine make a matrix of scatterplots?

[araichev]

Something similar to R's pairs or plotmatrix functions; see e.g. here?

[has2k1]

You cannot do that with plotnine, though it is a goal of the project that a third party library should be able to implement such compound plots using plotnine. However, you can do it with seaborn.

[marcio-pg]

Sorry for reviving this post Is this answer still valid in 2021?

[TyberiusPrime]

Still nothing 'out of the box' I believe.

Here's a start though:

import plotnine as p9, plotnine.data
import itertools


def plot_matrix(df, columns):
    pdf = []
    for a1, b1 in itertools.combinations(columns, 2):
        for (a,b) in ((a1, b1), (b1, a1)):
            sub = df[[a, b]].rename(columns={a: "x", b: "y"}).assign(a=a, b=b)
            pdf.append(sub)

    g = p9.ggplot(pd.concat(pdf))
    g += p9.geom_point(p9.aes('x','y'))
    g += p9.facet_grid('b~a', scales='free')
    return g#pd.concat(pdf)


plot_matrix(p9.data.economics, p9.data.economics.columns[1:])

[PaulHiemstra]

I was annoyed that the seaborn implementation of a pairplot did not allow me to change the shape of the point, only the color. This implementation in plotnine can do both shape and color.

Note that I use facet_wrap here and not facet_grid. This is to allow the use of scales='free'. It would be great if facet_grid allowed the scales to be free.

def get_data_for_column_combo(col1, col2, source, color_column, shape_column):
    '''
    Get the required data for a combination of two columns. Each call to this functions generates
    the data for one of the subplots in the pairplot. Color and shape data are appended as needed. 
    '''
    col_data = (source[[col1, col2]]
            .rename(columns={col1: 'values1', col2: 'values2'})
            .assign(col1=col1, col2=col2))
    if not color_column is None:
        col_data['color_column'] = source[color_column]
    if not shape_column is None:
        col_data['shape_column'] = source[shape_column]
    return col_data

def get_point_args(color_column, shape_column):
    '''
    Generate the appropriate input arguments to our geom_point. The names of the
    columns are fixed as these are generated in a standard way by `get_data_for_column_combo`. 
    But which should be included varies based on wheter or not color and shape are passed. 
    '''
    point_args = dict(x='values1', y='values2')

    if color_column is not None:
        point_args['color'] = 'color_column'
    if shape_column is not None:
        point_args['shape'] = 'factor(shape_column)'

    return point_args

def pairplot(source, columns, color_column=None, shape_column=None, use_facet_grid=False):
    '''
    This function creates a pairplot from the data in `source` based on the columns listed in `columns. 
    Optional arguments included passing a color and shape variabele, those will then determine the color and
    shape in the resulting pairplot. 

    By default we use `facet_wrap` as this allows us to use `scales='free'`. This is not how a pairplot
    usually works, so there is an option to force the use of `facet_grid` to get a more traditional plot.
    '''
    plot_data = pd.concat([
        get_data_for_column_combo(col1, col2, iris, color_column, shape_column) for col1, col2 in itertools.permutations(columns, 2)
    ])
    gg = ggplot(plot_data) + geom_point(aes(**get_point_args(color_column, shape_column)), alpha=0.4, size=2) 
    if use_facet_grid:
        return gg + facet_grid('col1 ~ col2')
    else:
        return gg + facet_wrap('~ col1 + col2', scales='free')

from plotnine import *
import pandas as pd
import itertools

pairplot(iris,['Sepal.Length', 'Sepal.Width', 'Petal.Length', 'Petal.Width'], color_column='Species', shape_column='cluster')

[PaulHiemstra]

@has2k1 Is there a third-party plotnine package where this function be sent to?

[b-lac]

@PaulHiemstra

I was annoyed that the seaborn implementation of a pairplot did not allow me to change the shape of the point, only the color. This implementation in plotnine can do both shape and color.

It looks like with seaborn you can change the markers in pairplot:

sns.pairplot(penguins, hue="species", markers=["o", "s", "D"])

[has2k1]

@PaulHiemstra, I don't think there is. Please open a separate issue to track this.

[PaulHiemstra]

@PaulHiemstra

I was annoyed that the seaborn implementation of a pairplot did not allow me to change the shape of the point, only the color. This implementation in plotnine can do both shape and color.

It looks like with seaborn you can change the markers in pairplot:

sns.pairplot(penguins, hue="species", markers=["o", "s", "D"])

The issue here is that the colors and the markers are based on the same variable, this is not what I was looking for. I want to base the color on one variable, and shape on the other. the case of my visualisation this enables me to inspect the difference between the fitted clusters of a Kmeans and the underlying Species of Iris.

[PaulHiemstra]

@PaulHiemstra, I don't think there is. Please open a separate issue to track this.

I'll do that.