Cluster Analysis: Tutorial with R

Jari Oksanen
January 26, 2014

Contents
1 Introduction
2 Hierarchic Clustering
2.1 Description of Classes . .
2.2 Numbers of Classes . . . .
2.3 Clustering and Ordination
2.4 Reordering a Dendrogram
2.5 Minimum Spanning Tree .
2.6 Cophenetic Distance . . .

.
.
.
.
.
.

1
4
4
5
6
7
8

3 Interpretation of Classes
3.1 Environmental Interpretation . . . . . . . . . . . . . . . . . . . .
3.2 Community Summaries . . . . . . . . . . . . . . . . . . . . . . .

8
9
10

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

4 Optimized Clustering at a Given Level

11
4.1 Optimum Number of Classes . . . . . . . . . . . . . . . . . . . . 11
5 Fuzzy Clustering

12

Introduction

In this tutorial we inspect classification. Classification and ordination are alternative strategies of simplifying data. Ordination tries to simplify data into a
map showing similarities among points. Classification simplifies data by putting
similar points into same class. The task of describing a high number of points
is simplified to an easier task of describing a low number of classes.

Hierarchic Clustering

The classification methods are available in standard R packages. The vegan

package does not have many support functions for classification, but we still
load vegan to have access to its data sets and some of its support functions.1
1 If you do not have a package, but get an error message, you must install package using
install.packages("vegan") or the installation menu.

A A
A
AA
B

B BB
B

A
B

A
B
B

B
B

B
B B

Figure 1: Distance between two clusters A and B defined by single, complete

and average linkage. Mark each of the linkage types in the connecting line. The
fusion level in the cluster dendrogram would be the length of the corresponding
connecting line of the linkage type.
R> library(vegan)
R> data(dune)
Hierarchic clustering (function hclust) is in standard R and available without loading any specific libraries. Hierarchic clustering needs dissimilarities as its
input. Standard R has function dist to calculate many dissimilarity functions,
but for community data we may prefer vegan function vegdist with ecologically
useful dissimilarity indices. The default index in vegdist is BrayCurtis:
R> d <- vegdist(dune)
Ecologically useful indices in vegan have an upper limit of 1 for absolutely
different sites (no shared species), and they are based on differences of abundances. In contrast, the standard Euclidean distance has no upper limit, but
varies with the sum of total abundances of compared sites when there are no
shared species, and uses squares of differences of abundances. There are many
other ecologically useful indices in vegdist, but BrayCurtis is usually not a
bad choice.
There are several alternative clustering methods in the standard function
hclust. We shall inspect three basic methods: single linkage, complete linkage
and average linkage. All these start in the same way: they fuse two most similar points to a cluster. They differ in the way they combine clusters to each
other, or new points to existing clusters (Fig. 1). In single linkage (a.k.a. nearest neighbour, or neighbour joining tree in genetics) the distance between two
clusters is the shortest possible distance among members of the clusters, or the
best of the friends. In complete linkage (a.k.a. furthest neighbour) the distance
between two clusters is the longest possible distance between the groups, or the
worst among the friends. In average linkage, the distance between the clusters

is the distance between cluster centroids. There are several alternative ways of
defining the average and defining the closeness, and hence a huge number of
average linkage methods. We only use one of these methods commonly known
as upgma. The lecture slides discuss the methods in more detail.
In the following we will compare three different clustering strategies. If you
want to plot three graphs side by side, you can divide the screen into three
panels by
R> par(mfrow=c(1,3))
This defines three panels side by side. You probably want to stretch the plotting
window if you are using this option. Alternatively, you can have three panels
above each other with
R> par(mfrow=c(3,1))
You can get back to the single panel mode with
R> par(mfrow=c(1,1))
You may also wish to use narrower empty margins for the panels:
R> par(mar=c(3,4,1,1)+.1)
The mar command defines plot margins in order bottom, left, up, right using
row height (text height) as a unit.
The single linkage clustering can be found with:
R> csin <- hclust(d, method="single")
R> csin
The dendrogram can be plotted with:
R> plot(csin)
The default is to plot an inverted tree with the root at the top, and branches
hanging down. You can force the branches down to the base line giving the hang
argument:
R> plot(csin, hang=-1)
If you plotted the csin tree twice you consumed two panels out of three you
have, and there will not be space for the next two trees in the same plot. In
that case you can start a new plot by issuing again the mfrow command and
then drawing csin again.
The complete linkage and average linkage methods are found in the same
way:
R>
R>
R>
R>

ccom <- hclust(d, method="complete")

plot(ccom, hang=-1)
caver <- hclust(d, method="aver")
plot(caver, hang=-1)

The vertical axes of the cluster dendrogram show the fusion level. The two
most similar observations are combined first, and they are at the same level
in all dendrograms. At the upper fusion levels, the scales diverge: they are
the shortest dissimilarities among cluster members in single linkage, the longest
possible dissimilarities in complete linkage, and the distances among cluster
centroids in average linkage (Fig. 1).
3

Figure 2: Vegemite is an Australian national delicacy made of yeast extract. The

vegemite function was named because its output is just as dense as Vegemite.

2.1

Description of Classes

One problem with hierarchic clustering is that it gives a classification of observations (plots, sampling units), but it does not tell how these classes differ
from each other. For community data, there is no information how the species
composition differs between classes (we return to this subject in Chapter 3.2).
The vegan package has function vegemite (Fig. 2) that can produce compact community tables ordered by a dendrogram, ordination or environmental
variables. With the help of these tables it is possible to see which species differ
in classification:
R> vegemite(dune, caver)
The vegemite command will always use one-character columns. If the observed values do not fit one character, the vegemite refuses to work. With argument scale you can recode the values to one-character width. The vegemite
has a graphical sister function tabasco that is described in section 2.4.

2.2

Numbers of Classes

The hierarchic clustering methods produce all possible levels of classifications.

The extremes are all observations in a single class, and each observation in
its private class. The user normally wants to have a clustering into a certain
number of classes. The fixed classification can be visually demonstrated with
rect.hclust function:
R>
R>
R>
R>

plot(csin, hang=-1)
rect.hclust(csin, 3)
plot(ccom, hang=-1)
rect.hclust(ccom, 3)

R> plot(caver, hang=-1)

R> rect.hclust(caver, 3)
Single linkage has a tendency to chain observations: most common case is
to fuse a single observation to an existing class: the single link is the nearest
neighbour, and a close neighbour is more probably in a large group than in a
small group or a lonely point. Complete linkage has a tendency to produce
compact bunches: complete link minimizes the spread within the cluster. The
average linkage is between these two extremes.
We can extract classification in a certain level using function cutree:
R> cl <- cutree(ccom, 3)
R> cl
This gives a numeric classification vector of cluster identities. The clusters are
numbered in the order the observations appear in the data: the first item will
always belong to cluster 1, and the numbering does not match the dendrogram.
We can tabulate the numbers of observations in each cluster:
R> table(cl)
We can compare two clustering schemes by cross-tabulation which gives as a
confusion matrix:
R> table(cl, cutree(csin, 3))
R> table(cl, cutree(caver, 3))
The confusion matrix tabulates the classifications against each other. The rows
give the first classification, and the columns the second classification. If the
classifications match and there is no confusion, each row and each column has
only one non-zero entry, but if the classes are divided between several classes in
the second classification, the row has several non-zero entries.

2.3

Clustering and Ordination

We can use ordination to display the observed dissimilarities among points.

A natural choice is to use metric scaling a.k.a. principal coordinates analysis
(PCoA) that maps observed dissimilarities linearly onto low-dimensional graph
using the same dissimilarities we had in our clustering.
The metric scaling can be performed with standard R function cmdscale:
R> ord <- cmdscale(d)
We can display the results using vegan function ordiplot that can plot
results of any vegan ordination function and many non-vegan ordination functions, such as cmdscale, prcomp and princomp (the latter for principal components analysis):
R> ordiplot(ord)
We got a warning because ordiplot tries to plot both species and sites in the
same graph, and the cmdscale result has no species scores. We do not need to
care about this warning.
There are many vegan functions to overlay classification onto ordination.
For distinct, non-overlapping classes convex hulls are practical:
5

Figure 3: A dendrogram is similar to a mobile: branches can turn around, but

the mobile is the same. Alexander Calder: Red Mobile.
R> ordihull(ord, cl, lty=3)
R> ordispider(ord, cl, col="blue", label=TRUE)
R> ordiellipse(ord, cl, col="red")
For overlapping classes we can use ordispider. If we are not interested in
individual points, but on class centroids and average dispersions, we can use
ordiellipse, both with the same arguments as ordihull.
The other clustering results can be seen with:
R>
R>
R>
R>

ordiplot(ord, dis="si")
ordihull(ord, cutree(caver, 3))
ordiplot(ord, dis="si")
ordicluster(ord, csin)

We set here explicitly the display argument to display = "sites" to avoid the
annoying and useless warnings. The contrasting clustering strategies (nearest
vs. furthest vs. average neighbour) are evident in the shapes of the clusters.
Single linkage clusters are chained, complete linkage clusters are compact, and
average linkage clusters between these two.
The vegan package has a special function to display the cluster fusions
in ordination. The ordicluster function combines sites and cluster centroids
similarly as the average linkage method:
R> ordiplot(ord, dis="si")
R> ordicluster(ord, caver)
We can prune the top level fusions to highlight the clustering:
R> ordiplot(ord, dis="si")
R> ordicluster(ord, caver, prune=2)

2.4

Reordering a Dendrogram

The leaves of a dendrogram do not have a natural order. You can take a branch
and turn around its root, and the tree is the same (see Fig. 3).
R has two alternative dendrogram presentations: the hclust result object
and a general dendrogram object. The cluster type can be changed with:
R> den <- as.dendrogram(caver)
The dendrograms are more general, and several methods are available for their
manipulation and analysis. It is possible to re-order the leaves of a dendrogram
so that they match as closely as possible an external variable.
6

In the following we rearrange the dendrogram so that the ordering of leaves

corresponds as closely as possible with the first ordination axis:
R> x <- scores(ord, display = "sites", choices = 1)
R> oden <- reorder(den, x)
We plot the dendrorgrams together:
R>
R>
R>
R>

par(mfrow=c(2,1))
plot(den)
plot(oden)
par(mfrow=c(1,1))

The reordered dendrogram may also give more regularly structured community
table:
R> vegemite(dune, oden)
The vegemite function has a graphical sister function tabasco2 that can also
display the dendrogram. Moreover, it defaults to rerrange the dendrogram by
the first axis of Correspondence Analysis:
R> tabasco(dune, caver)
Correspondence Analysis packs similar species next to each other, and similar
sites next to each other and gives a good diagonal representation of the data. If
you want to see the original ordering of the sample plots, you must set Rowv =
FALSE:
R> tabasco(dune, caver, Rowv = FALSE)
R> tabasco(dune, oden, Rowv = FALSE)
tabasco function also defaults to order species to match the ordering of sites
unless you set Colv = FALSE.

2.5

Minimum Spanning Tree

In the mathematical graph theory, tree is a connected graph without loops,

spanning tree is a tree connecting all points, and minimum spanning tree is the
shortest of such trees. Minimum spanning tree (mst) is closely related to single
linkage clustering, which also connects all points with minimum total connecting
distance. However, mst really combines points, whereas R representations of
single linkage clustering hierarchically connects clusters instead of single points.
mst can be found with vegan function spantree3
R> mst <- spantree(d)
mst can be overlaid onto an ordination with lines command:
R> ordiplot(ord, dis="si")
R> lines(mst, ord)
2 So

called because it is similar to vegemite but hotter.

are many other implementations MST in R, but the implementation in vegan
probably is the fastest.
3 There

Alternatively, the tree can be displayed with a plot command that tries to find a
locally optimal configuration for points to reproduce the distances among points
along the tree. Internally the function uses Sammon scaling (sammon function
in the MASS package) to find the configuration of points. Sammon scaling is
a variant of metric scaling trying to reproduce relative distances among points
and it is optimal for showing the local (vs. global) structure of points.
R> plot(mst, type="t")

2.6

Cophenetic Distance

The estimated distance between two points is the level at which they are fused in
the dendrogram, or the height of the root. A good clustering method correctly
reproduces the actual dissimilarities. The distance estimated from a dendrogram is called cophenetic distance. The name echoes the origins of hierarchic
clustering in old fashioned numeric taxonomy. Standard R function cophenetic
estimates the distances among all points from a dendrogram.
We can visually inspect the cophenetic distances against observed dissimilarities. In the following, abline adds a line with zero intercept and slope of
one, or the equivalence line. We also set equal scaling on x and y axes (asp =
1):
R>
R>
R>
R>
R>
R>

plot(d, cophenetic(csin), asp=1)

abline(0, 1)
plot(d, cophenetic(ccom), asp=1)
abline(0, 1)
plot(d, cophenetic(caver), asp=1)
abline(0, 1)

In single linkage clustering, the cophenetic distances are as long as or shorter

than the observed distances: the distance between groups is the shortest possible
distance between its members. In complete linkage clustering, the cophenetic
distances are as long as or longer than observed distances: the distance between
two groups is the longest possible distance between groups. In average linkage
clustering, the cophenetic distance is the average of observed distances (cf. Fig.
1).
The correlation between observed and cophenetic distances is called the
cophenetic correlation:
R> cor(d, cophenetic(csin))
R> cor(d, cophenetic(ccom))
R> cor(d, cophenetic(caver))
The ranking of these cophenetic correlations is not entirely random: it is guaranteed that average linkage (upgma) method maximizes the cophenetic correlation.

Interpretation of Classes

We commonly want to use classes for prediction of external variables this

is the idea om Finnish forest type system, EU Water Framework Directive and
8

theory of gradient analysis. We make classes from community composition and

the classes should describe community composition, and we assume the classes
predict environmental conditions.

3.1

Environmental Interpretation

For interpreting the classes we take into use the the environmental data:
R> data(dune.env)
The cutree function produced a numerical classification vector, but we need a
factor of classes:
R> cl <- factor(cl)
We can visually display the differences in environmental conditions using boxplot.
The only continuous variable we have is the thickness of the A1 horizon, but we
can change the Moisture class to a numeric variable:
R> Moist <- with(dune.env, as.numeric(as.character(Moisture)))
Compare the result of this to a simpler looking alternative:
R> with(dune.env, as.numeric(Moisture))
The latter uses the internal representation (1, 2, 3, 4) of the factor instead of
its nominal values (1, 2, 4, 5) that we can get with the added as.character
command.
The boxplot is produced with:
R> boxplot(Moist ~ cl, notch=TRUE)
The boxplot shows the data extremes (whiskers and the points for extreme
cases), lower and upper quartiles and the median. The notches show the approximate 95 % confidence limits of the median: if the notches do not overlap,
the medians probably are different at level P < 0.05. For small data sets like
this, the confidence limits are wide, and you get a warning of waist going over
the shoulders.
You can use anova for rigid numerical testing of differences among classes.
In R, anova is an analysis method for linear models (lm):
R> anova(lm(Moist ~ cl))
This performs parametric anova test. If we want to use non-parametric permutation tests with the same test statistic, we can use redundancy analysis:
R> anova(rda(Moist ~

cl))

This works nicely in this case because the y-variate (Moist) is in the current
workspace: the command would not work for variates in data frames unless we
attach the data frame or use the command as with(varechem, ...).
We can use confusion matrix to compare factor variables and classification:
R> with(dune.env, table(cl, Management))

3.2

Community Summaries

We can use the labdsv package for analysing community classification.4

R> library(labdsv)
The package has several functions for summarizing community composition in
classes. The species occurrence frequencies in classes can be found with
R> const(dune, cl)
and the mean abundances with
R> importance(dune, cl)
The labdsv has function indval for indicator species analysis.5 The indicator species analysis combines frequency tables (const) and mean abundance
tables (importance) and finds species that are significantly concentrated into
specific classes. The model is fitted with:
R> mod <- indval(dune, as.numeric(cl))
The indval result object contains the following items:
R> names(mod)
The item relfrq is the same as the output of const, and relabu is based on
the output of importance, but each row is normalized to unit sum so that
the table shows the distribution of total abundances in classes. These values
are combined into indicator value (indval) of each species for each class, and
maxcls gives the class to which species has the highest indicator value (indcls).
Finally, vector pval gives the probability of finding higher indicator value in
random permutations, and if this probability is low, the species is a significant
indicator.
Each item can be inspected separately:
R> mod$maxcls
R> mod$pval
Often it is more practical to look at the summary. The "short" summary lists the
significant indicator species with some extra information:
R> summary(mod)
The "long" summary lists all species with their indicator values for classes, but
replaces non-significant values with a dot:
R> summary(mod, type = "long")
4 If the following command fails, you must install labdsv using install.packages("labdsv")
or from the installation menu.
5 In old labdsv versions the function was called duleg, but the usage was similar as for
indval below.

10

Optimized Clustering at a Given Level

Hierarchic clustering has a history: it starts from the bottom, and combines
observations to clusters. At higher levels it can only fuse existing clusters to
others, but it cannot break old groups to build better ones. Often we are most
interested in classification at a low number of clusters, and these have a long
historic burden. Once we have decided upon the number of classes, we can often
improve the classification at a given number of clusters. A tool for this task is
K-means clustering that produces an optimal solution for a given number of
classes.
K-means clustering has one huge limitation for community data: it works
in Euclidean metric which rarely is meaningful for community data. In hierarchic clustering we could use non-Euclidean Bray-Curtis dissimilarity, but we do
not have this choice here. However, we can work around some problems with
appropriate standardization. Hellinger transformation has been suggested to be
very useful with Euclidean metric and community data: square root of data
divided by row (site) totals. The Hellinger transformation can be performed
with vegan function decostand which has many other useful standardizations
and transformations:
R> ckm <- kmeans(decostand(dune, "hell"), 3)
R> ckm$cluster
We can display the results in the usual way:
R> ordiplot(ord, dis="si")
R> ordihull(ord, ckm$cluster, col="red")

4.1

Optimum Number of Classes

The K-means clustering optimizes for the given number of classes, but what is
the correct number of classes? Arguably there is no correct number, because
the World is not made of distinct classes, but all classifications are patriarchal
artifacts. However, some classifications may be less useful (and worse) than
others.
Function cascadeKM in vegan finds some of these criteria for a sequence of Kmeans clusterings. The default method is the CalinskiHarabasz criterion which
is the F ratio of anova for the predictor variables (species in community data).
The test uses Euclidean metric and we use again the Hellinger transformation
(if you try without transformation or different transformations, you can see that
transformations with decostand have capricious effects: obviously the World is
not classy in this case).
R> ccas <- cascadeKM(decostand(dune, "hell"), 2, 15)
R> plot(ccas, sortq=TRUE)
We gave the smallest (2) and largest (15) number of inspected classes in the
call, and sorted the plot by classification. Each class is shown by a different
colour, and bends in the vertical colour bars show non-hierarchical clustering
where the upper levels are no simple fusions of lower levels. The highest value of
the criterion shows the optimal number of classes. This kind of graph is called

11

Figure 4: Piet Mondrian: Composition in red, blue and yellow.

Mondrian plot after an excellent Dutch-born artist Piet Mondriaan of De Stijl
school (Fig. 4).6

Fuzzy Clustering

We have so far worked with classification methods which implicitly assume that
there are distinct classes. The World is not made like that. If there are classes,
they are vague and have intermediate and untypical cases. With one word, they
are fuzzy.
Fuzzy classification means that each observation has a certain probability of
belonging to a class. In the crisp case, it has probability 1 of belonging to one
class, and probability 0 of belonging to any other class. In a fuzzy case, it has
probability < 1 for the best class, and probabilities > 0 for several other classes
(and the sum of these probabilities is 1).
Fuzzy classification is similar to K-means clustering in finding the optimal
classification for a given number of classes, but the produced classification is
fuzzy: the result is a probability profile of class membership.
The fuzzy clustering is provided by function fanny in package cluster7 .
Fuzzy clustering defaults to Euclidean metric, but current versions can accept
any dissimilarities. In the following, we also need to use lower membership
exponent (memb.exp): the default 2 gives complete fuzziness and fails here, and
lower values give crisper classifications.8
R> library(cluster)
R> cfuz <- fanny(d, 3, memb.exp=1.7)
The function returns an object with the following items:
R> names(cfuz)
6 A similar plot for cross classifications is called a Marimekko plot: Google to see what it
looks like and how it is constructed.
7 All functions in the cluster package have stupid names.
8 Old version of fanny may not have all these options and can fail; you should upgrade R
to get a new version of the cluster package.

12

Here membership is the probability profile of belonging to a certain class, and

clustering is the most likely crisp classification.
It is difficult to show the fuzzy results graphically, but here is one idea:
R>
R>
R>
R>

ordiplot(ord, dis="si")
ordiplot(ord, dis="si", type="n")
stars(cfuz$membership, locatio=ord, draw.segm=TRUE, add=TRUE, scale=FALSE, len=0.1)
ordihull(ord, cfuz$clustering, col="blue")

This uses stars function (with many optional parameters) to show the probability profile, and draws a convex hull of the crisp classification. The size of the
sector shows the probability of the class membership, and in clear cases, one of
the segments is dominant.

13