In OllieFord/ChangepointVis: Interactive Visualization for Change Point Exploration and Labeling

Tidying k-means clustering

knitr::opts_chunk$set(message = FALSE, warning = FALSE)

K-means clustering serves as a very useful example of tidy data, and especially the distinction between the three tidying functions: tidy, augment, and glance.

Let's start by generating some random two-dimensional data with three clusters. Data in each cluster will come from a multivariate gaussian distribution, with different means for each cluster:

library(dplyr)
library(ggplot2)
library(purrr)
library(tibble)
library(tidyr)

set.seed(27)

centers <- tibble(
  cluster = factor(1:3), 
  num_points = c(100, 150, 50),  # number points in each cluster
  x1 = c(5, 0, -3),              # x1 coordinate of cluster center
  x2 = c(-1, 1, -2)              # x2 coordinate of cluster center
)

labelled_points <- centers %>%
  mutate(
    x1 = map2(num_points, x1, rnorm),
    x2 = map2(num_points, x2, rnorm)
  ) %>% 
  select(-num_points) %>% 
  unnest(x1, x2)

ggplot(labelled_points, aes(x1, x2, color = cluster)) +
  geom_point()

This is an ideal case for k-means clustering. We'll use the built-in kmeans function, which accepts a dataframe with all numeric columns as it's primary argument.

points <- labelled_points %>% 
  select(-cluster)

kclust <- kmeans(points, centers = 3)
kclust
summary(kclust)

The output is a list of vectors, where each component has a different length. There's one of length r nrow(points): the same as our original dataset. There are a number of elements of length 3: withinss, tot.withinss, and betweenss- and centers is a matrix with 3 rows. And then there are the elements of length 1: totss, tot.withinss, betweenss, and iter.

These differing lengths have a deeper meaning when we want to tidy our dataset: they signify that each type of component communicates a different kind of information.

• cluster (r nrow(points) values) contains information about each point
• centers, withinss and size (3 values) contain information about each cluster
• totss, tot.withinss, betweenss, and iter (1 value) contain information about the full clustering

Which of these do we want to extract? There is no right answer: each of them may be interesting to an analyst. Because they communicate entirely different information (not to mention there's no straightforward way to combine them), they are extracted by separate functions. augment adds the point classifications to the original dataset:

library(broom)

augment(kclust, points)

The tidy function summarizes on a per-cluster level:

tidy(kclust)

And as it always does, the glance function extracts a single-row summary:

glance(kclust)

broom and dplyr for exploratory clustering

While these summaries are useful, they would not have been too difficult to extract out from the dataset yourself. The real power comes from combining these analyses with dplyr.

Let's say we want to explore the effect of different choices of k, from 1 to 9, on this clustering. First cluster the data 9 times, each using a different value of k, then create columns containing the tidied, glanced and augmented data:

kclusts <- tibble(k = 1:9) %>%
  mutate(
    kclust = map(k, ~kmeans(points, .x)),
    tidied = map(kclust, tidy),
    glanced = map(kclust, glance),
    augmented = map(kclust, augment, points)
  )

kclusts

We can turn these into three separate datasets each representing a different type of data: Then tidy the clusterings three ways: using tidy, using augment, and using glance. Each of these goes into a separate dataset as they represent different types of data.

clusters <- kclusts %>%
  unnest(tidied)

assignments <- kclusts %>% 
  unnest(augmented)

clusterings <- kclusts %>%
  unnest(glanced, .drop = TRUE)

Now we can plot the original points, with each point colored according to the predicted cluster.

p1 <- ggplot(assignments, aes(x1, x2)) +
  geom_point(aes(color = .cluster)) + 
  facet_wrap(~ k)
p1

Already we get a good sense of the proper number of clusters (3), and how the k-means algorithm functions when k is too high or too low. We can then add the centers of the cluster using the data from tidy:

p2 <- p1 + geom_point(data = clusters, size = 10, shape = "x")
p2

The data from glance fits a different but equally important purpose: it lets you view trends of some summary statistics across values of k. Of particular interest is the total within sum of squares, saved in the tot.withinss column.

ggplot(clusterings, aes(k, tot.withinss)) +
  geom_line()

This represents the variance within the clusters. It decreases as k increases, but one can notice a bend (or "elbow") right at k=3. This bend indicates that additional clusters beyond the third have little value. (See here for a more mathematically rigorous interpretation and implementation of this method). Thus, all three methods of tidying data provided by broom are useful for summarizing clustering output.

OllieFord/ChangepointVis documentation built on Aug. 26, 2019, 7:53 p.m.