Distributions of multiple numerical features with Seaborn

When given a set of numerical features, it is desirable to plot all of them using for example violinplots, to get a sense of their respective distributions. Seaborn can do this automatically using the violinplot function, but this doesn't work so well when the features have widely different ranges.



In [ ]:

    
import string

import pandas as pd
import numpy as np
import seaborn as sns

Get some random data

Create a function that returns a data frame where each feature is a random numerical variable where we can control the max amplitude.



In [ ]:

    
def get_random_numerical_data(size, *amplitudes):
    n = len(amplitudes)
    data = np.random.random((size, n)) * np.array(amplitudes).reshape(1, n)
    return pd.DataFrame(data=data, columns=pd.Series(list(string.ascii_uppercase[:n]), name="feature"))



In [ ]:

    
get_random_numerical_data(5, 1, 2)



In [ ]:

    
get_random_numerical_data(500, 1, 2, 3, 4).describe().loc[['count', 'std', 'max']]

Plotting all features directly with Seaborn

Works well when the data range is small



In [ ]:

    
df_small_range = get_random_numerical_data(500, 1, 2, 3, 4)
sns.violinplot(df_small_range)



In [ ]:

    
df_big_range = get_random_numerical_data(500, 1, 10, 100, 1000)
sns.violinplot(df_big_range)

Changing the y-scale to log doesn't help much



In [ ]:

    
df_big_range = get_random_numerical_data(500, 1, 10, 100, 1000)
h = sns.violinplot(df_big_range)
h.set_yscale('log')

Plotting distributions on separate figures using Matplotlib

We could standardize all columns with Scikit Learn but then we use all sense of scale. Ideally, we want to plot each data on a separate plot. We can do this with a little matplotlib function.



In [ ]:

    
import matplotlib.pyplot as plt
def featureplot(df, nrows=1, ncols=1, figsize=(12,8), plotfunc=sns.violinplot):
    """Plot the dataframe features"""
    width, height = figsize
    fig, axes = plt.subplots(nrows, ncols, figsize=(width, height * nrows));
    i = 0
    plots_per_figure = max(df.shape[1] // (nrows * ncols), 1)
    if nrows == 1 and ncols == 1:
        axes = [axes]
    if nrows > 1 and ncols > 1:
        axes = chain.from_iterable(axes)  # flatten the nested list
    for j, ax in zip(range(plots_per_figure, df.shape[1] + 1, plots_per_figure), axes):
        plotfunc(df.iloc[:, i:j], ax=ax)
        i = j
    plt.tight_layout()



In [ ]:

    
featureplot(df_big_range, ncols=4)

Plotting on separate columns using Seaborn only

However we can do this directly with Seaborn using a FacetGrid if we put the data in long form.



In [ ]:

    
df_big_range_lf = df_big_range.stack().reset_index(name="value").drop('level_0', axis=1)#.reset_index()  # don't keep the index
df_big_range_lf.head()



In [ ]:

    
# size is the height of each figure and aspect is the with/height aspect ratio of each figure.
sns.FacetGrid(df_big_range_lf, col="feature", hue="feature",
              sharey=False, size=7, aspect=8/12.0/2.0).map(sns.violinplot, "value", orient="v")

Appendix: stack vs pivot

Pivot can be implemented by using set_index and unstack.



In [ ]:

    
test = pd.DataFrame({'foo':["one"] * 3 + ["two"] * 3, 'bar': list("ABC")*2, 'baz': list(range(6))})



In [ ]:

    
test



In [ ]:

    
test.pivot('foo', 'bar', 'baz')



In [ ]:

    
test.set_index(['foo','bar']).unstack()['baz']