## Pretty Joint Plots

In this tutorial, I’ll be using Python to create a neat, customizable joint plot–– inspired by the jointplot graphics found in Seaborn. Joint plots are great for plotting bivariate datasets, as they’re readily legible and provide high information content. Here, I’ll be creating a KDE joint plot with a modern color palette but a.simple design that resembles an old-school topographical map.

To follow along, you’ll need a working knowledge of statistics as well as NumPy and SciPy in order to generate a dataset; matplotlib to create the figure itself; and Seaborn to enhance matplotlib’s somewhat bland aesthetics. Note: I’m also using mpld3 to generate an interactive plot, but this is completely optional.

## Package Installation

Assuming you’re familiar with pip, we can further simplify installation by installing conda (Although I highly recommend using Anaconda’s python distribution, which comes with conda, NumPy, SciPy and matplotlib pre-installed):

Once conda is installed, getting NumPy, SciPy, and matplotlib is a easy as:

conda will automatically install the packages and their requirements using pre-compiled binaries, so installation times should be on the order of seconds instead of tens of minutes.

Seaborn is only available through pip, however:

Once these packages have been installed, you can get started!

## Data Generation

If you don’t already have some data lying around to plot, you can easily create some using a bivariate normal distribution with NumPy. First, we’ll need an easy way to generate pseudo-random positive-semidefinite matrices for our distribution’s covariances:

Then, we can define a pair of means and covariance matrices for our distribution:

Here, we’ve generated a pair of 2-D means from $\mu \sim I*X + Y$, where $I$ is a random integer pulled from $U(1,10)$, and $X$ and $Y$ are both random float arrays pulled from $N(0,1)$. If this all seems cryptic–– don’t worry! These parameters are completely arbitrary, and the point is just that we’re generating an interesting distribution to plot.

To sample 500 points from our bivariate distribution, we can use numpy.random.multivariate_normal:

So, now we have some data to plot!

## Scatter Plot

Before we go on to making a KDE of our data, let’s just make a joint plot with a scatter plot in the main grid and marginal histograms. First let’s import Seaborn and set our plot style to 'dark'

We can then set the x- and y-axes limits:

Although, for aesthetic purposes, we should probably adjust these values so that the all the points are clearly visible. I find that scaling up each axis by a factor of $1.5$ works fairly well:

These values will be useful for making the KDE joint plot as well.

Now, we can use the following code to generate the figure:

## Kernel Density Estimation

There are many ways to create a kernel density estimator in Python, but in this tutorial we’ll be using scipy.stats.gaussian_kde:

To extract densities from the gaussian_kde objects, we need to create some meshgrids using our max- and min-values from before:

## Contour Plot

Now that we have KDEs for both the marginals and the bivariate distribution, we can make our final plot with the following code: