Using Self-Organizing Maps with SOMbrero to cluster a numeric dataset

To limit this documentation size, most figures are not displayed in the version of the vignette included in the package. To see all figures, you can rerun the vignettes, changing the following option to TRUE:

knitr::opts_chunk$set(include = FALSE)

Alternatively, this compilation is also available at: http://sombrero.nathalievialaneix.eu/articles/c-doc-numericSOM.html

Basic package description

SOMbrero implements different variants of the Self-Organizing Map algorithm (also called Kohonen's algorithm). To run the standard version of the algorithm, use the function trainSOM() on a data frame or matrix with numerical columns. The standard numeric SOM and its use in SOMbrero are illustrated below.

This documentation only considers the case of numerical data.

library("ggplot2")
library("SOMbrero")

Arguments

The trainSOM function has several arguments, but only the first one is required. This argument is x.data which is the dataset used to train the SOM. In this documentation, it is passed to the function as a matrix or a data frame with numerical variables in columns and observations of these variables in rows.

The other arguments are the same as the arguments passed to the initSOM function (they are parameters defining the algorithm, see help(initSOM) for further details).

Outputs

The trainSOM function returns an object of class somRes (see help(trainSOM) for further details on this class).

Graphics

The following table indicates which graphics are available for a numeric SOM.

| What

SOM or SC
Type | SOM
Energy


| Obs



| Prototypes



| Add



| SuperCluster
(no what)


| Obs



| Prototypes



| Add



| |:------------|:--------|:-----|:------------|:-----|:-------------|:-----|:------------|:-----| | (no type) | x | | | | | | | | | hitmap | | x | | | | x | | | | color | | x | x | x | | x | x | | | lines | | x | x | x | | x | x | x | | meanline | | x | | x | | x | | x | | barplot | | x | x | x | | x | x | x | | pie | | | | x | | | | x | | boxplot | | x | | x | | x | | x | | 3d | | | x | | | | | | | poly.dist | | | x | | | | x | | | umatrix | | | x | | | | | | | smooth.dist | | | x | | | | | | | mds | | | x | | | | x | | | grid.dist | | | x | | | | | | | words | | | | x | | | | | | names | | x | | x | | | | | | grid | | | | | | | x | | | dendrogram | | | | | x | | | | | dendro3d | | | | | x | | | |

First case study: simulated data in $[0,1]^2$

The first case study shows the clustering of points randomly distributed in the square $[0,1]^2$. The data are generated by:

set.seed(4031719)
the.data <- data.frame("x1" = runif(500), "x2" = runif(500))
ggplot(the.data, aes(x = x1, y = x2)) + geom_point() + theme_bw()

Training the SOM

The numeric SOM algorithm is used to cluster the data:

set.seed(1105)
# run the SOM algorithm with 10 intermediate backups and 2000 iterations
my.som <- trainSOM(x.data=the.data, dimension=c(5,5), nb.save=10, maxit=2000, 
                   scaling="none", radius.type="letremy", topo="square",
                   dist.type = "letremy")

The energy evolves as described in the following graphic:

plot(my.som, what="energy")

Clustering

The resulting clustering distribution can be visualized by the hitmap:

plot(my.som, what = "obs", type = "hitmap")

The observations are almost uniformly distributed on the map.

The clustering component allows us to plot the initial data according to the final clustering.

# prepare a vector of colors
my.colors <- rainbow(prod(my.som$parameters$the.grid$dim))[my.som$clustering]

# points depicted with the same color are in the same final cluster
plot(my.som$data[,1], my.som$data[,2], col=my.colors, pch=19, xlab="x1", 
     ylab="x2", main="Data according to final clustering")

Clustering interpretation

The values of the prototypes can be represented with the plot function and help interpret the clusters:

par(mfrow=c(1,2))
plot(my.som, what="prototypes", type="color", var=1)
plot(my.som, what="prototypes", type="color", var=2)

Here, the interpretation is simple enough: high values of the first variables x1 are located at the top of the map and small values at the bottom of the map. Large values of x2 are located at the right hand side of the map, whereas, small values are located at the left hand side.

We obtain the same results with a similar plot on the observation mean values:

par(mfrow=c(1,2))
plot(my.som, what="obs", type="color", var=1)
plot(my.som, what="obs", type="color", var=2)

The prototypes coordinates are also registered for each intermediate backup so they can be displayed on different graphics to see the evolution in the prototype organization.

# Get the neighbours between prototypes
values <- protoDist(my.som, "neighbors")
tmp <- data.frame("prot1" = rep.int(1:prod(my.som$parameters$the.grid$dim), 
                                    times=sapply(values, length)), 
                  "nei" = as.numeric(as.character(names(unlist(values)))))
tmp <- tmp[tmp[ ,1] < tmp[ ,2], ]

# plot the prototypes
par(mfrow=c(2, 5),mar=c(3,2,2,1))
invisible(sapply(1:my.som$parameters$nb.save, function(ind){
  plot(my.som$backup$prototypes[[ind]][,1], my.som$backup$prototypes[[ind]][,2],
       xlab="", ylab="", main=c("iteration ", my.som$backup$steps[ind]))
  for (i in 1:nrow(tmp)){
    segments(x0=my.som$backup$prototypes[[ind]][tmp[i,1],1], 
             y0=my.som$backup$prototypes[[ind]][tmp[i,1],2],
             x1=my.som$backup$prototypes[[ind]][tmp[i,2],1], 
             y1=my.som$backup$prototypes[[ind]][tmp[i,2],2], 
             col="red", pch=19)
  }
}))

At the beginning of the algorithm, the prototypes are randomly distributed in [0,1]^2 and then, they organize as a regular rectangular grid in $[0,1]^2$.

Second case study: the iris dataset

This second case study is performed on the famous (Fisher's or Anderson's) iris data set that gives the measurements in centimeters of the variables sepal length and width and petal length and width, respectively, for 50 flowers from each of 3 species of iris (setosa, versicolor, and virginica).

Training the SOM

NB: In the following analysis, variables are centered and scaled to unit variance, which is the default behavior of the algorithm.

The first four variables of the data set (that are the numeric variables) are used to map each flower on the SOM grid.

set.seed(255)
# run the SOM algorithm with verbose set to TRUE
iris.som <- trainSOM(x.data = iris[,1:4], dimension = c(5,5), verbose = TRUE, 
                     nb.save = 5, topo = "hexagonal")
iris.som

As the energy is registered during the intermediate backups, we can have a look at its evolution.

plot(iris.som, what="energy")

Here the energy does not stabilize as in the case of dist.type="letremy" because the Gaussian annealing of the neighborhood is continuous and not stepwise.

Resulting clustering

The clustering component contains the final classification of the dataset. It is a vector with length equal to the number of rows of the input dataset.

iris.som$clustering
table(iris.som$clustering)

which can also be visualized by a hitmap plot:

plot(iris.som, what="obs", type="hitmap")

To assess the relevance of each explanatory variable in the definition of the clusters, the function summary includes an ANOVA with the predictor being the clustering, for each (numeric) input variable.

summary(iris.som)

Here, all variables have significantly different means among the different clusters and can thus be considered to be relevant for the clustering definition.

Another useful function is predict.somRes. This function predicts the neuron to which a new observation would be assigned. The first argument must be a somRes object and the second one the new observation. Let us have a try on the first observation of the iris data set:

# call predict.somRes
predict(iris.som, iris[1,1:4])
# check the result of the final clustering with the SOM algorithm
iris.som$clustering[1]

Clustering interpretation

Graphics common to observations and prototypes

Some graphics are shared between observations and prototypes and can be used to display the prototypes' or the observations' values for the different variables in the neurons of the map. For observations, the mean values are sometimes displayed instead of the individual values.

par(mfrow = c(2,2))
plot(iris.som, what = "obs", type = "color", variable = 1)
plot(iris.som, what = "obs", type = "color", variable = 2)
plot(iris.som, what = "obs", type = "color", variable = 3)
plot(iris.som, what = "obs", type = "color", variable = 4)
plot(iris.som, what = "prototypes", type = "lines", show.names = TRUE) + 
  theme(axis.text.x = element_blank())
plot(iris.som, what = "obs", type = "barplot", show.names = TRUE) + 
  theme(axis.text.x = element_blank())

Some neurons are empty (no observations affected to them): neurons 3, 7, 8, 9, 10, 13, 14, 15, 19, 20, 24. They are the illustration of a large difference between observations classified in clusters 4-5 with the rest of the observations (to some extend, the same can be said about observations in cluster 25).

Clusters 4-5 are characterized by the following facts (visible on "color" plots): larger values for Sepal.Width, smaller values for Petal.Length and Petal.Width and average values for Sepal.Length. Similar conclusions are obtained from the "lines" and "barplot" plots. Please note that, for these two last plots, it is advised to display the neuron numbers (option show.names = TRUE) because the hexagonal display is not properly rendered in them.

More graphics on observations

Individual information on observations in clustered can be obtained with the following plots:

plot(iris.som, what = "obs", type = "boxplot", show.names = TRUE)
plot(iris.som, what = "obs", type = "lines", show.names = TRUE)
plot(iris.som, what = "obs", type = "names", show.names = TRUE)

They display either the observation distribution within the cluster for all the variables in the dataset (for "boxplot" and "lines") or the names (row numbers by default) of the observations classified within the cluster.

More graphics on prototypes

Some more graphics handling prototypes have been implemented:

par(mfrow=c(2,2))
plot(iris.som, what = "prototypes", type = "3d", variable = 1)
plot(iris.som, what = "prototypes", type = "3d", variable = 2)
plot(iris.som, what = "prototypes", type = "3d", variable = 3)
plot(iris.som, what = "prototypes", type = "3d", variable = 4)

Also, some graphics are provided to visualize the distance between prototypes on the grid:

plot(iris.som, what = "prototypes", type = "poly.dist", show.names = FALSE)
plot(iris.som, what = "prototypes", type = "umatrix")
plot(iris.som, what = "prototypes", type = "smooth.dist")
plot(iris.som, what = "prototypes", type = "mds")
plot(iris.som, what = "prototypes", type = "grid.dist")

These graphics show that there is a big gap (large distances) between the top left corner and the rest of the map and between the top right corner and the rest of the map (which is consistent with what was already observed in the previous plots). In addition, the bottom right corner of the maps has a few neurons (21, 22, 23) that are very closed to each others.

Graphics showing an additional variable

Additional factor

The clustering can be analyzed together with an additional variable (here, the flower species) using what = "add":

class(iris$Species)
levels(iris$Species)
plot(iris.som, what = "add", type = "pie", variable = iris$Species) +
  scale_fill_brewer(type = "qual") + 
  guides(fill = guide_legend(title = "Species"))

This plot shows that the clustering produced by the SOM is indeed relevant to identify the three different species of iris: they are well separated on the map and almost all clusters only contain one species of iris. The Setosa species is the most distinct from the other two, isolated in the top left corner of the map.

Additional numerical vector

The "color" plot available for "add" is similar to the "obs" or "prototypes" cases. Here we choose the first variable of the iris data set as an additional variable to illustrate its use. We thus obtain the same plot as above (see section Graphics common to observations and prototypes).

plot(iris.som, what = "add", type = "color", variable = iris$Sepal.Length, 
     show.names = FALSE)
Additional numerical matrix or data frame

The "lines", "barplot", "radar" and "boxplot" plots available for "add" are similar to the "obs" or "prototypes" cases.

"words" is only implemented for an additional variable. In this case, the additional variable must be a contingency matrix: the words used on the plot are the names of the columns and the presence or lack of the word is expressed by respectively 1 or 0. The size of the words on the grid depends on the rate of presence in the observations of the current neuron. To illustrate its use, we define a contingency table my.cont.mat that corresponds to the flower Species:

my.cont.mat <- matrix(data=c(rep(c(rep(1,50), rep(0,150)), 2), rep(1,50)), 
                      nrow = 150, ncol = 3)
colnames(my.cont.mat) <- levels(iris$Species)
head(my.cont.mat)
plot(iris.som, what = "add", type = "words", variable = my.cont.mat, 
     show.names = FALSE)
Additional non-numerical vector

"names" is similar to the "names" case implemented for "obs". Here we choose to give the argument variable the row names of the iris data set: so we obtain the same plot as above (see More graphics on observations).

plot(iris.som, what = "add", type = "names", variable = rownames(iris)) 

Similarly, this plot can be used with the variable iris$Species:

plot(iris.som, what = "add", type = "names", variable = iris$Species)

which gives exactly the same plot as before for type "words" with the contingency matrix corresponding to the variable iris$Species.

Analyze the projection quality

quality(iris.som)
qualities <- quality(iris.som)

By default, the quality function calculates both quantization and topographic errors. It is also possible to specify which one you want using the argument quality.type.

The topographic error value varies between 0 (good projection quality) and 1 (poor projection quality). Here, the topographic quality of the mapping is equal to r round(qualities$topographic, 2), which means that around r round(qualities$topographic * 100, 1)% of the observations have a second best unit in the neighborhood of the best matching unit.

The quantization error is an unbounded positive number. The closer it is to 0, the better the projection quality.

Building super classes from the resulting SOM

In the SOM algorithm, the number of clusters is necessarily close to the number of neurons on the grid (not necessarily equal as some neurons may have no observations assigned to them). This - quite large - number may not suit the original data for a clustering purpose.

A usual way to address clustering with SOM is to perform a hierarchical clustering on the prototypes. This clustering is directly available in the package SOMbrero using the function superClass. To do so, you can first have a quick overview to decide on the number of super clusters which suits your data.

plot(superClass(iris.som))

By default, the function plots both a dendrogram and the evolution of the percentage of explained variance. Here, 3 super clusters seem to be the best choice. The output of superClass is a somSC class object. Basic functions have been defined for this class:

my.sc <- superClass(iris.som, k = 3)
summary(my.sc)
plot(my.sc, plot.var = FALSE)

Like plot.somRes, the function plot.somSC has an argument type which offers many different plots and can thus be combined with most of the graphics produced by plot.somSC:

plot(my.sc, type = "grid")
plot(my.sc, type = "dendro3d")

Case "grid" fills the grid with colors according to the super clustering (and can provide a legend). Case "dendro3d" plots a 3d dendrogram.

A couple of plots from plot.somRes are also available for the super clustering. Some identify the super clusters with colors:

plot(my.sc, what = "obs", type = "hitmap", maxsize = 20)
plot(my.sc, what = "prototypes", type = "lines")
plot(my.sc, what = "prototypes", type = "barplot")
plot(my.sc, what = "prototypes", type = "mds")

And some others identify the super clusters with titles:

plot(my.sc, what = "prototypes", type = "color", variable = "Sepal.Length")
plot(my.sc, what = "prototypes", type = "poly.dist")

It is also possible to consider an additional variable using the argument what='add':

plot(my.sc, what = "add", type = "pie", variable = iris$Species) +
  scale_fill_brewer(type = "qual")
plot(my.sc, what = "add", type = "color", variable = iris$Sepal.Length)

Super cluster number 2 is located at the top left hand corner of the map and associated with the Setosa species. SC 3 is in the opposite corner and associated mainly with Virginica whereas SC 1 is in the diagonal between these two corners and associated with Versicolor.

Session information

This vignette has been computed with the following environment:

sessionInfo()


Try the SOMbrero package in your browser

Any scripts or data that you put into this service are public.

SOMbrero documentation built on Jan. 4, 2022, 1:07 a.m.