clouds

clouds is a package that implements non-parametric clustering algorithims for groups of numeric data. A cloud is a numerical vector with a weight vector associated from which an empircal cumulative distribution function (ECDF) can be computed.

clouds defines a pseudo-distance function to compare ECDFs. So given a set of clouds -- e.g. an experiment with a response variable and treatment groups, where each group will define a cloud -- a distance matrix for the clouds can be computed. Based on this matrix we can apply different well-known clustering methods to group the clouds.

Currently there are two methods implemented k_clouds which uses k-means for the clustering and j_clouds which uses Ward's hierarchichal algorithm as implemented in stats::hclust() with method = "ward.D".

Installation

Install from GitHub with devtools:

devtools::install_github("nebulae-co/clouds")

Usage

Lets make some data from five different normal distributions:

library("dplyr")

n_distributions <- 5L
n_groups <- 4L
group_size <- 12
N <- group_size * n_groups
distributions <- letters[1:n_distributions]
groups <- LETTERS[1:(n_distributions * n_groups)]

mean <- 1:5
sd <- c(2, 1, 0.5, 1, 2)

sample <- data_frame(
  y = c(mapply(rnorm, mean = mean, sd = sd, MoreArgs = list(n = N))),
  distribution = rep(distributions, each = N),
  group = rep(groups, each = group_size))

sample

#> # A tibble: 240 x 3
#>             y distribution group
#>         <dbl>        <chr> <chr>
#>  1  2.0355661            a     A
#>  2  2.4251103            a     A
#>  3  0.7001281            a     A
#>  4 -0.6477998            a     A
#>  5  1.2728224            a     A
#>  6  1.6601561            a     A
#>  7  1.2103591            a     A
#>  8 -1.0395949            a     A
#>  9  1.7951571            a     A
#> 10  3.5448991            a     A
#> # ... with 230 more rows

Now lets load clouds and get the k_clouds clusters:

library("clouds")

sample %>%
  k_clouds(var = "y", group = "group", k = 5, threshold = 0.01)

#>    group cluster
#> 1      A       e
#> 2      B       e
#> 3      C       e
#> 4      D       e
#> 5      E       e
#> 6      F       e
#> 7      G       c
#> 8      H       e
#> 9      I       c
#> 10     J       c
#> 11     K       c
#> 12     L       c
#> 13     M       d
#> 14     N       b
#> 15     O       b
#> 16     P       d
#> 17     Q       a
#> 18     R       a
#> 19     S       a
#> 20     T       a

Compare with the j_clouds clusters:

sample %>%
  j_clouds(var = "y", group = "group", k = 5)

#>    group cluster
#> 1      A       a
#> 2      B       a
#> 3      C       a
#> 4      D       a
#> 5      E       b
#> 6      F       b
#> 7      G       b
#> 8      H       b
#> 9      I       c
#> 10     J       c
#> 11     K       c
#> 12     L       c
#> 13     M       d
#> 14     N       d
#> 15     O       d
#> 16     P       d
#> 17     Q       e
#> 18     R       e
#> 19     S       e
#> 20     T       e

About

This is mostly result of Julian's research for his Master Thesis.

Daniel has done the packaging.

Still a work in progress.