Multivariate Visualization¶
- Used to compare univariate distributions between groups
- Used to find relations between 2 variables
- Used to see how a relationship changes between 2 variables across groups
Plots¶
- Best at representing two variables on the X and Y axes
- Three variables possible with 3D plots but should be avoided if possible.
- Third variable usually represented as color, shape, or another plot
- Most suitable for comparison of relationship or distributions across groups
In [82]:
# Setup the libraries
%matplotlib inline
import seaborn as sns
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
sns.set(color_codes=True)
# lets load the data again
weather_df = pd.read_csv("https://raw.githubusercontent.com/vega/vega-datasets/gh-pages/data/weather.csv")
cars_df = pd.read_json("https://github.com/vega/vega-datasets/raw/gh-pages/data/cars.json")
Plotting Two Variables¶
- Both matplot and seaborn could be used
- Scatter plot typically used
- Can help detect relations
- Time searies plots are a special form of these plots showing relationship to time
- line plots possible with time series
In [5]:
# Scatter plot for Acceleration X Miles per gallon
plt.scatter(cars_df.Acceleration, cars_df.Miles_per_Gallon)
Out[5]:
In [17]:
# Scatter plot with Seaborn
sns.lmplot(x='Acceleration', y='Miles_per_Gallon', data=cars_df, fit_reg=False)
Out[17]:
In [18]:
# Seaborn can also fit a regression line to show the direction of the relation
sns.lmplot(x='Acceleration', y='Miles_per_Gallon', data=cars_df)
Out[18]:
Plotting Three or More Variables¶
- This typically means we are comparing subgroups in our data
- You can use colors (hue) to represent different groups
- You can also plot different groups in different plot side by side
- Organized in rows, or columns
- The groups are typically categorical variables
- Consider Subdividing continuous variables if you want to use as groups
In [19]:
# Plotting 3 variables, using jue
sns.lmplot(x='Acceleration', y='Miles_per_Gallon', hue='Cylinders',data=cars_df, fit_reg=False)
Out[19]:
In [69]:
# Possible on matplotlib as well
# Showing legend properly is not easy
plt.scatter(cars_df.Acceleration, cars_df.Miles_per_Gallon, c=cars_df.Cylinders)
plt.legend()
Out[69]:
In [67]:
# Same thing with matplotlib
# Notice the legend is missing
colors = {
2:'r',
3:'g',
4:'b',
5:'y',
6:'w',
8:'k',
}
# We draw each cylinder plot separatly
for x in sorted(set(cars_df.Cylinders)):
d = cars_df[cars_df.Cylinders == x]
plt.scatter(d.Acceleration, d.Miles_per_Gallon, c=colors.get(x), label=x)
plt.legend(title="Cylinders")
Out[67]:
Tip¶
Use seaborn whenever possible
In [72]:
# See what happens when I replace hue with col
sns.lmplot(x='Acceleration', y='Miles_per_Gallon', col='Cylinders',data=cars_df, fit_reg=False)
Out[72]:
In [73]:
# now row
sns.lmplot(x='Acceleration', y='Miles_per_Gallon', row='Cylinders',data=cars_df, fit_reg=False)
Out[73]:
Using row/col In Seaborn Plots¶
- Avoid using it with variables that have many values
- Will create many plots
- Difficult to compare
- Use when variable has few values
In [119]:
# Let's examine relationship between wind and percipitation
g = sns.lmplot(x='wind', y='precipitation', data=weather_df, fit_reg=False)
In [120]:
# Distribution of percipitation
sns.distplot(weather_df.precipitation)
# The values are bunched up close to zero
Out[120]:
In [121]:
# Using log scale on the y access will make the plot clearer
g = sns.lmplot(x='wind', y='precipitation', data=weather_df, fit_reg=False)
g.set(yscale="log")
# No clear relationship
Out[121]:
In [123]:
# Let's see if it is the same in all locations
g = sns.lmplot(x='wind', y='precipitation', hue='location', data=weather_df, fit_reg=False)
g.set(yscale="log")
# Hue not making comparison easy
Out[123]:
In [124]:
# two plots are better for comparison
g = sns.lmplot(x='wind', y='precipitation', col='location', data=weather_df, fit_reg=False)
g.set(yscale="log")
# Still no clear relation
Out[124]:
In [125]:
# Let's also break it based on weather types
g = sns.lmplot(x='wind', y='precipitation', col='location', row='weather', data=weather_df, fit_reg=False)
g.set(yscale="log")
Out[125]:
EDA¶
- Remember it is an investigation
- Sometimes our investigation takes as to a dead end
- We think of different ways to break our data
- Possibly rescale our axes, like the log scale
- Think about missing data that we might get
- There might be nothing interesting in the relationship
- Examine other relationships
Pair Plots¶
- Use to get a quick overview of the numeric data that you have
- Diagonal represents the distributions
- Of diagonals gives you relationships between the variables
- Use to find insights that you can dig deeper into
In [2]:
sns.pairplot(weather_df)
Out[2]:
Remember The Univariate Plots in Seaborn?¶
- Violin, stip, swarm, count, and dist plots
- You can use them for multivariate comparison of distributions
- You can set both the y and hue to partition your data
- Cannot place them in rows and columns
- Use them with factor plots
In [3]:
# Try to plot 2 different plots of each (10 plot in total)
# where you set the y and then the hue to see how the
# plots will behave differently
Factor Plots¶
- Use it to further breakdown the distributional plots
- Allows you to place them in rows and columns as well
- Examine documentation for factor plot and attempt to plot 2 different distributional plots in rows or columns for comparison
In [126]:
# Your turn to show off FactorPlot