I have a data set with huge number of features, so analysing the correlation matrix has become very difficult. I want to plot a correlation matrix which we get using dataframe.corr() function from pandas library. Is there any built-in function provided by the pandas library to plot this matrix?
This question is related to
python
pandas
matplotlib
data-visualization
information-visualization
The answer is
Form correlation matrix, in my case zdf is the dataframe which i need perform correlation matrix.
corrMatrix =zdf.corr()
corrMatrix.to_csv('sm_zscaled_correlation_matrix.csv');
html = corrMatrix.style.background_gradient(cmap='RdBu').set_precision(2).render()
# Writing the output to a html file.
with open('test.html', 'w') as f:
print('<!DOCTYPE html><html lang="en"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-widthinitial-scale=1.0"><title>Document</title></head><style>table{word-break: break-all;}</style><body>' + html+'</body></html>', file=f)
Then we can take screenshot. or convert html to an image file.
Surprised to see no one mentioned more capable, interactive and easier to use alternatives.
A) You can use plotly:
Just two lines and you get:
interactivity,
smooth scale,
colors based on whole dataframe instead of individual columns,
column names & row indices on axes,
zooming in,
panning,
built-in one-click ability to save it as a PNG format,
auto-scaling,
comparison on hovering,
bubbles showing values so heatmap still looks good and you can see values wherever you want:
import plotly.express as px
fig = px.imshow(df.corr())
fig.show()
B) You can also use Bokeh:
All the same functionality with a tad much hassle. But still worth it if you do not want to opt-in for plotly and still want all these things:
from bokeh.plotting import figure, show, output_notebook
from bokeh.models import ColumnDataSource, LinearColorMapper
from bokeh.transform import transform
output_notebook()
colors = ['#d7191c', '#fdae61', '#ffffbf', '#a6d96a', '#1a9641']
TOOLS = "hover,save,pan,box_zoom,reset,wheel_zoom"
data = df.corr().stack().rename("value").reset_index()
p = figure(x_range=list(df.columns), y_range=list(df.index), tools=TOOLS, toolbar_location='below',
tooltips=[('Row, Column', '@level_0 x @level_1'), ('value', '@value')], height = 500, width = 500)
p.rect(x="level_1", y="level_0", width=1, height=1,
source=data,
fill_color={'field': 'value', 'transform': LinearColorMapper(palette=colors, low=data.value.min(), high=data.value.max())},
line_color=None)
color_bar = ColorBar(color_mapper=LinearColorMapper(palette=colors, low=data.value.min(), high=data.value.max()), major_label_text_font_size="7px",
ticker=BasicTicker(desired_num_ticks=len(colors)),
formatter=PrintfTickFormatter(format="%f"),
label_standoff=6, border_line_color=None, location=(0, 0))
p.add_layout(color_bar, 'right')
show(p)
You can use imshow() method from matplotlib
import pandas as pd
import matplotlib.pyplot as plt
plt.style.use('ggplot')
plt.imshow(X.corr(), cmap=plt.cm.Reds, interpolation='nearest')
plt.colorbar()
tick_marks = [i for i in range(len(X.columns))]
plt.xticks(tick_marks, X.columns, rotation='vertical')
plt.yticks(tick_marks, X.columns)
plt.show()
Try this function, which also displays variable names for the correlation matrix:
def plot_corr(df,size=10):
'''Function plots a graphical correlation matrix for each pair of columns in the dataframe.
Input:
df: pandas DataFrame
size: vertical and horizontal size of the plot'''
corr = df.corr()
fig, ax = plt.subplots(figsize=(size, size))
ax.matshow(corr)
plt.xticks(range(len(corr.columns)), corr.columns);
plt.yticks(range(len(corr.columns)), corr.columns);
Seaborn's heatmap version:
import seaborn as sns
corr = dataframe.corr()
sns.heatmap(corr,
xticklabels=corr.columns.values,
yticklabels=corr.columns.values)
If your main goal is to visualize the correlation matrix, rather than creating a plot per se, the convenient pandas styling options is a viable built-in solution:
import pandas as pd
import numpy as np
rs = np.random.RandomState(0)
df = pd.DataFrame(rs.rand(10, 10))
corr = df.corr()
corr.style.background_gradient(cmap='coolwarm')
# 'RdBu_r' & 'BrBG' are other good diverging colormaps
Note that this needs to be in a backend that supports rendering HTML, such as the JupyterLab Notebook. (The automatic light text on dark backgrounds is from an existing PR and not the latest released version, pandas 0.23).
Styling
You can easily limit the digit precision:
corr.style.background_gradient(cmap='coolwarm').set_precision(2)
Or get rid of the digits altogether if you prefer the matrix without annotations:
corr.style.background_gradient(cmap='coolwarm').set_properties(**{'font-size': '0pt'})
The styling documentation also includes instructions of more advanced styles, such as how to change the display of the cell the mouse pointer is hovering over. To save the output you could return the HTML by appending the render() method and then write it to a file (or just take a screenshot for less formal purposes).
Time comparison
In my testing, style.background_gradient() was 4x faster than plt.matshow() and 120x faster than sns.heatmap() with a 10x10 matrix. Unfortunately it doesn't scale as well as plt.matshow(): the two take about the same time for a 100x100 matrix, and plt.matshow() is 10x faster for a 1000x1000 matrix.
Saving
There are a few possible ways to save the stylized dataframe:
- Return the HTML by appending the
render()method and then write the output to a file. - Save as an
.xslxfile with conditional formatting by appending theto_excel()method. - Combine with imgkit to save a bitmap
- Take a screenshot (for less formal purposes).
Update for pandas >= 0.24
By setting axis=None, it is now possible to compute the colors based on the entire matrix rather than per column or per row:
corr.style.background_gradient(cmap='coolwarm', axis=None)
statmodels graphics also gives a nice view of correlation matrix
import statsmodels.api as sm
import matplotlib.pyplot as plt
corr = dataframe.corr()
sm.graphics.plot_corr(corr, xnames=list(corr.columns))
plt.show()
If you dataframe is df you can simply use:
import matplotlib.pyplot as plt
import seaborn as sns
plt.figure(figsize=(15, 10))
sns.heatmap(df.corr(), annot=True)
Along with other methods it is also good to have pairplot which will give scatter plot for all the cases-
import pandas as pd
import numpy as np
import seaborn as sns
rs = np.random.RandomState(0)
df = pd.DataFrame(rs.rand(10, 10))
sns.pairplot(df)
You can observe the relation between features either by drawing a heat map from seaborn or scatter matrix from pandas.
Scatter Matrix:
pd.scatter_matrix(dataframe, alpha = 0.3, figsize = (14,8), diagonal = 'kde');
If you want to visualize each feature's skewness as well - use seaborn pairplots.
sns.pairplot(dataframe)
Sns Heatmap:
import seaborn as sns
f, ax = pl.subplots(figsize=(10, 8))
corr = dataframe.corr()
sns.heatmap(corr, mask=np.zeros_like(corr, dtype=np.bool), cmap=sns.diverging_palette(220, 10, as_cmap=True),
square=True, ax=ax)
The output will be a correlation map of the features. i.e. see the below example.
The correlation between grocery and detergents is high. Similarly:
Pdoducts With High Correlation:- Grocery and Detergents.
- Milk and Grocery
- Milk and Detergents_Paper
- Milk and Deli
- Frozen and Fresh.
- Frozen and Deli.
From Pairplots: You can observe same set of relations from pairplots or scatter matrix. But from these we can say that whether the data is normally distributed or not.
Note: The above is same graph taken from the data, which is used to draw heatmap.
Source: Stackoverflow.com