The cassowaryr package provides functions to compute scagnostics on
pairs of numeric variables in a data set.
The term scagnostics refers to scatter plot diagnostics, originally described by John and Paul Tukey. This is a collection of techniques for automatically extracting interesting visual features from pairs of variables. This package is an implementation of graph theoretic scagnostics developed by Wilkinson, Anand, and Grossman (2005) in pure R.
remotes::install_github("numbats/cassowaryr")data("anscombe_tidy")
sc_outlying(anscombe$x1, anscombe$y1)
sc_outlying(anscombe$x2, anscombe$y2)
sc_outlying(anscombe$x3, anscombe$y3)
sc_outlying(anscombe$x4, anscombe$y4)CAlculate Scagnostics on Scatterplots Over Wads of Associated Real numberYs in R
A 2-d scatter plot can be represented by a combination of three graphs which are computed directly from the Delauney-Voroni tesselation.
- A minimum spanning tree weighted by the lengths of the Delauney triangles
- The convex hull of the points i.e. the outer segments of the triangulation
- The alpha hull (also called concave hull) i.e. formed by connect the outer edges of triangles that are enclosed within a ball of radius alpha.
All graph based scagnostic measures are computed with respect to these three graphs.
Prior to graph construction decisions must be made about filtering
outliers (done with respect to the distribution of edge lengths in the
triangulation) and thinning the size of the graphs by performing binning
(for computational speed). For the moment we can forge ahead without
these but it is worth keeping in mind that the package needs to be
flexible enough to include them.
There are also opportunities to experiment with the preprocessing here.
Two MST measures “clumpy” and “outlying” are known to cause problems.
As all the graph based measures rely on the triangulation, they could be computed lazily. More concretely, if you are only interested in computing “skinny” you don’t need to compute the spanning tree. If you computed “skinny” but wanted to then compute “convex” you shouldn’t need to reconstruct the alpha-hull and so on… To begin let’s not worry about that and focus on implementations of each measure.
These are computed directly from the 2-d point clouds, and do not need to be constructed from the graph.