The R Project SVN R

% File src/library/stats/man/hclust.Rd
% Part of the R package, https://www.R-project.org
% Copyright 1995-2026 R Core Team
% Distributed under GPL 2 or later

\name{hclust}
\title{Hierarchical Clustering}
\alias{hclust}
\alias{plot.hclust}
\alias{print.hclust}
\description{
  Hierarchical cluster analysis on a set of dissimilarities and
  methods for analyzing it.
}
\usage{
hclust(d, method = "complete", members = NULL)

\method{plot}{hclust}(x, labels = NULL, hang = 0.1, check = TRUE,
     axes = TRUE, frame.plot = FALSE, ann = TRUE,
     main = "Cluster Dendrogram",
     sub = NULL, xlab = NULL, ylab = "Height", \dots)
}
\arguments{
  \item{d}{a dissimilarity structure as produced by \code{dist}.}

  \item{method}{the agglomeration method to be used.  This should
    be (an unambiguous abbreviation of) one of
    \code{"ward.D"}, \code{"ward.D2"}, \code{"single"}, \code{"complete"},
    \code{"average"} (= \abbr{UPGMA}), \code{"mcquitty"} (= \abbr{WPGMA}),
    \code{"median"} (= \abbr{WPGMC}) or \code{"centroid"} (= \abbr{UPGMC}).}

  \item{members}{\code{NULL} or a vector with length size of
    \code{d}. See the \sQuote{Details} section.}

  \item{x}{an object of the type produced by \code{hclust}.}

  \item{hang}{The fraction of the plot height by which labels should hang
    below the rest of the plot.
    A negative value will cause the labels to hang down from 0.}
  \item{check}{logical indicating if the \code{x} object should be
    checked for validity.  This check is not necessary when \code{x}
    is known to be valid such as when it is the direct result of
    \code{hclust()}.  The default is \code{check=TRUE}, as invalid
    inputs may crash \R due to memory violation in the internal C
    plotting code.}
  \item{labels}{A character vector of labels for the leaves of the
    tree.  By default the row names or row numbers of the original data are
    used.  If \code{labels = FALSE} no labels at all are plotted.}

  \item{axes, frame.plot, ann}{logical flags as in \code{\link{plot.default}}.}

  \item{main, sub, xlab, ylab}{character strings for
    \code{\link{title}}.  \code{sub} and \code{xlab} have a non-NULL
    default when there's a \code{tree$call}.}
  \item{\dots}{Further graphical arguments.  E.g., \code{cex} controls
    the size of the labels (if plotted) in the same way as \code{\link{text}}.}
}
\value{
  An object of class \code{"hclust"} which describes the
  tree produced by the clustering process.
  The object is a list with components:
  \item{merge}{an \eqn{n-1} by 2 matrix.
    Row \eqn{i} of \code{merge} describes the merging of clusters
    at step \eqn{i} of the clustering.
    If an element \eqn{j} in the row is negative,
    then observation \eqn{-j} was merged at this stage.
    If \eqn{j} is positive then the merge
    was with the cluster formed at the (earlier) stage \eqn{j}
    of the algorithm.
    Thus negative entries in \code{merge} indicate agglomerations
    of singletons, and positive entries indicate agglomerations
    of non-singletons.}

  \item{height}{a set of \eqn{n-1} real values (non-decreasing for
    ultrametric trees).
    The clustering \emph{height}: that is, the value of
    the criterion associated with the clustering
    \code{method} for the particular agglomeration.}

  \item{order}{a vector giving the permutation of the original
    observations suitable for plotting, in the sense that a cluster
    plot using this ordering and matrix \code{merge} will not have
    crossings of the branches.}

  \item{labels}{labels for each of the objects being clustered.}

  \item{call}{the call which produced the result.}

  \item{method}{the cluster method that has been used.}

  \item{dist.method}{the distance that has been used to create \code{d}
    (only returned if the distance object has a \code{"method"}
    attribute).}

  There are \code{\link{print}}, \code{\link{plot}} and \code{identify}
  (see \code{\link{identify.hclust}}) methods and the
  \code{\link{rect.hclust}()} function for \code{hclust} objects.
}
\details{
  This function performs a hierarchical cluster analysis
  using a set of dissimilarities for the \eqn{n} objects being
  clustered.  Initially, each object is assigned to its own
  cluster and then the algorithm proceeds iteratively,
  at each stage joining the two most similar clusters,
  continuing until there is just a single cluster.
  At each stage distances between clusters are recomputed
  by the Lance--Williams dissimilarity update formula
  according to the particular clustering method being used.

  A number of different clustering methods are provided.  \emph{Ward's}
  minimum variance method aims at finding compact, spherical clusters.
  The \emph{complete linkage} method finds similar clusters. The
  \emph{single linkage} method (which is closely related to the minimal
  spanning tree) adopts a \sQuote{friends of friends} clustering
  strategy.  The other methods can be regarded as aiming for clusters
  with characteristics somewhere between the single and complete link
  methods.  Note however, that methods \code{"median"} and
  \code{"centroid"} are \emph{not} leading to a \emph{monotone distance}
  measure, or equivalently the resulting dendrograms can have so called
  \emph{inversions} or \emph{reversals} which are hard to interpret,
  but note the trichotomies in \bibcitet{R:Legendre+Legendre:2012}.

  Two different algorithms are found in the literature for Ward
  clustering.  The one used by option \code{"ward.D"} (equivalent to the
  only Ward option \code{"ward"} in \R versions \eqn{\le} 3.0.3)
  \emph{does not} implement the clustering criterion of
  \bibcitet{R:Ward:1963}, whereas option \code{"ward.D2"} implements
  that criterion \bibcitep{R:Murtagh+Legendre:2014}.  With the latter, the
  dissimilarities are \emph{squared} before cluster updating, and the
  clustering corresponds to \code{"ward.D"} with squared distances, see
  also the example.
  Note that \code{\link[cluster]{agnes}(*, method="ward")} corresponds
  to \code{hclust(*, "ward.D2")}.

  If \code{members != NULL}, then \code{d} is taken to be a
  dissimilarity matrix between clusters instead of dissimilarities
  between singletons and \code{members} gives the number of observations
  per cluster.  This way the hierarchical cluster algorithm can be
  \sQuote{started in the middle of the dendrogram}, e.g., in order to
  reconstruct the part of the tree above a cut (see examples).
  Dissimilarities between clusters can be efficiently computed (i.e.,
  without \code{hclust} itself) only for a limited number of
  distance/linkage combinations, the simplest one being \emph{squared}
  Euclidean distance and centroid linkage.  In this case the
  dissimilarities between the clusters are the squared Euclidean
  distances between cluster means.

  In hierarchical cluster displays, a decision is needed at each merge to
  specify which subtree should go on the left and which on the right.
  Since, for \eqn{n} observations there are \eqn{n-1} merges,
  there are \eqn{2^{(n-1)}} possible orderings for the leaves
  in a cluster tree, or dendrogram.
  The algorithm used in \code{hclust} is to order the subtree so that
  the tighter cluster is on the left (the last, i.e., most recent,
  merge of the left subtree is at a lower value than the last
  merge of the right subtree).
  Single observations are the tightest clusters possible,
  and merges involving two observations place them in order by their
  observation sequence number.
}
\note{
  Method \code{"centroid"} has been recommended to be used with
  \emph{squared} Euclidean distances \code{dist(*)^2}, i.e., clustering
  differently than the traditional \abbr{UPGMC}, where other authors
  strongly prefer the original.
}
\references{
  \bibinfo{R:Becker+Chambers+Wilks:1988}{note}{(S version)}
  \bibinfo{R:Murtagh:1985}{note}{(For algorithmic details of algorithms used)}
  \bibshow{*,
    R:Anderberg:1973,
    R:Becker+Chambers+Wilks:1988,
    R:Everitt:1974,
    R:Gordon:1999,
    R:Hartigan:1975,
    R:Legendre+Legendre:2012,
    R:McQuitty:1966,
    R:Murtagh:1985,
    R:Sneath+Sokal:1973,
    R:Ward:1963}
}
\author{
  The \code{hclust} function is based on Fortran code
  contributed to STATLIB by F. Murtagh.
}
\seealso{
  \code{\link{identify.hclust}}, \code{\link{rect.hclust}},
  \code{\link{cutree}}, \code{\link{dendrogram}}, \code{\link{kmeans}}.

  For the Lance--Williams formula and methods that apply it generally,
  see \code{\link[cluster]{agnes}} from package \CRANpkg{cluster}.
}
\examples{
require(graphics)

### Example 1: Violent crime rates by US state

hc <- hclust(dist(USArrests), "ave")
plot(hc)
plot(hc, hang = -1)

## Do the same with centroid clustering and *squared* Euclidean distance,
## cut the tree into ten clusters and reconstruct the upper part of the
## tree from the cluster centers.
hc <- hclust(dist(USArrests)^2, "cen")
memb <- cutree(hc, k = 10)
cent <- matrix(numeric(), 10, 4)
for(k in 1:10)
  cent[k,] <- colMeans(USArrests[memb == k, , drop = FALSE])

hc1 <- hclust(dist(cent)^2, method = "cen", members = table(memb))
opar <- par(mfrow = c(1, 2))
plot(hc,  labels = FALSE, hang = -1, main = "Original Tree")
plot(hc1, labels = FALSE, hang = -1, main = "Re-start from 10 clusters")
par(opar)

### Example 2: Straight-line distances among 10 US cities
##  Compare the results of algorithms "ward.D" and "ward.D2"

mds2 <- -cmdscale(UScitiesD)
plot(mds2, type="n", axes=FALSE, ann=FALSE)
text(mds2, labels=rownames(mds2), xpd = NA)

hcity <- list(
    D.o = hclust(UScitiesD, "ward.D"), # "wrong"
    D2o = hclust(UScitiesD, "ward.D2"),# similar to next two
    D   = hclust(1/2* (UScitiesD)^2,  "ward.D"),  # Ward(1963) ESS  based
    D2  = hclust(sqrt(1/2)*UScitiesD, "ward.D2")) #       sqrt(ESS) based

opar <- par(mfrow = c(2, 2), mgp = c(1.2, .3, 0), mar = .1 + c(3,4,2,1))
 plot(hcity$D.o, hang=-1, col=2) # different clustering than the others
 plot(hcity$D2o, hang=-1) # [1,2]
 plot(hcity$D,   hang=-1) # [2,1]
 plot(hcity$D2,  hang=-1) # [2,2] as above, [1,2], just y-axis numbers differ
par(opar)
mrg <- sapply(hcity, `[[`, "merge") # defining cluster tree (apart from height)
ht  <- sapply(hcity, `[[`, "height")
stopifnot(exprs = {
   any(mrg[,2] != mrg[,1])
       mrg[,2] == mrg[,3:4] # all eq., but the first
   all.equal(ht[,"D2" ], sqrt(ht[,"D"]))
   all.equal(ht[,"D2"], sqrt(1/2)*ht[,"D2o"])
})

}
\keyword{multivariate}
\keyword{cluster}