| Title: | Fast Hierarchical Clustering Routines for R and 'Python' |
|---|---|
| Description: | This is a two-in-one package which provides interfaces to both R and 'Python'. It implements fast hierarchical, agglomerative clustering routines. Part of the functionality is designed as drop-in replacement for existing routines: linkage() in the 'SciPy' package 'scipy.cluster.hierarchy', hclust() in R's 'stats' package, and the 'flashClust' package. It provides the same functionality with the benefit of a much faster implementation. Moreover, there are memory-saving routines for clustering of vector data, which go beyond what the existing packages provide. For information on how to install the 'Python' files, see the file INSTALL in the source distribution. Based on the present package, Christoph Dalitz also wrote a pure 'C++' interface to 'fastcluster': <https://lionel.kr.hs-niederrhein.de/~dalitz/data/hclust/>. |
| Authors: | Daniel Müllner [aut, cph, cre], Google Inc. [cph] |
| Maintainer: | Daniel Müllner <[email protected]> |
| License: | FreeBSD | GPL-2 | file LICENSE |
| Version: | 1.2.6 |
| Built: | 2024-11-08 06:15:44 UTC |
| Source: | CRAN |
The fastcluster package provides efficient algorithms for hierarchical, agglomerative clustering. In addition to the R interface, there is also a Python interface to the underlying C++ library, to be found in the source distribution.
The function hclust provides clustering when the
input is a dissimilarity matrix. A dissimilarity matrix can be
computed from vector data by dist. The
hclust function can be used as a drop-in replacement for
existing routines: stats::hclust and
flashClust::hclust alias
flashClust::flashClust. Once the
fastcluster library is loaded at the beginning of the code, every
program that uses hierarchical clustering can benefit immediately and
effortlessly from the performance gain
When the package is loaded, it overwrites the function
hclust with the new code.
The function hclust.vector provides memory-saving routines
when the input is vector data.
Further information:
R documentation pages: hclust,
hclust.vector
A comprehensive User's manual:
fastcluster.pdf. Get this from the R
command line with vignette('fastcluster').
JSS paper: doi:10.18637/jss.v053.i09.
See the author's home page for a performance comparison: https://danifold.net/fastcluster.html.
Daniel Müllner
https://danifold.net/fastcluster.html
# Taken and modified from stats::hclust # # hclust(...) # new method # hclust.vector(...) # new method # stats::hclust(...) # old method require(fastcluster) require(graphics) 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.vector(USArrests, "cen") # squared Euclidean distances hc$height <- hc$height^2 memb <- cutree(hc, k = 10) cent <- NULL for(k in 1:10){ cent <- rbind(cent, colMeans(USArrests[memb == k, , drop = FALSE])) } hc1 <- hclust.vector(cent, method = "cen", members = table(memb)) # squared Euclidean distances hc1$height <- hc1$height^2 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)# Taken and modified from stats::hclust # # hclust(...) # new method # hclust.vector(...) # new method # stats::hclust(...) # old method require(fastcluster) require(graphics) 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.vector(USArrests, "cen") # squared Euclidean distances hc$height <- hc$height^2 memb <- cutree(hc, k = 10) cent <- NULL for(k in 1:10){ cent <- rbind(cent, colMeans(USArrests[memb == k, , drop = FALSE])) } hc1 <- hclust.vector(cent, method = "cen", members = table(memb)) # squared Euclidean distances hc1$height <- hc1$height^2 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)
This function implements hierarchical clustering with the same interface as hclust from the stats package but with much faster algorithms.
hclust(d, method="complete", members=NULL)hclust(d, method="complete", members=NULL)
d |
a dissimilarity structure as produced by |
method |
the agglomeration method to be used. This must be (an
unambiguous abbreviation of) one of |
members |
|
See the documentation of the original function
hclust in the stats package.
A comprehensive User's manual
fastcluster.pdf is available as a vignette. Get this from the R command line with vignette('fastcluster').
An object of class 'hclust'. It encodes a stepwise dendrogram.
Daniel Müllner
https://danifold.net/fastcluster.html
fastcluster, hclust.vector, stats::hclust
# Taken and modified from stats::hclust # # hclust(...) # new method # stats::hclust(...) # old method require(fastcluster) require(graphics) 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 <- NULL for(k in 1:10){ cent <- rbind(cent, 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)# Taken and modified from stats::hclust # # hclust(...) # new method # stats::hclust(...) # old method require(fastcluster) require(graphics) 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 <- NULL for(k in 1:10){ cent <- rbind(cent, 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)
This function implements hierarchical, agglomerative clustering with memory-saving algorithms.
hclust.vector(X, method="single", members=NULL, metric='euclidean', p=NULL)hclust.vector(X, method="single", members=NULL, metric='euclidean', p=NULL)
X |
an |
method |
the agglomeration method to be used. This must be (an
unambiguous abbreviation of) one of |
members |
|
metric |
the distance measure to be used. This must be one of
|
p |
parameter for the Minkowski metric. |
The function hclust.vector provides clustering when the
input is vector data. It uses memory-saving algorithms which allow
processing of larger data sets than hclust does.
The "ward", "centroid" and "median" methods
require metric="euclidean" and cluster the data set with
respect to Euclidean distances.
For "single" linkage clustering, any dissimilarity
measure may be chosen. Currently, the same metrics are implemented as the
dist function provides.
The call
hclust.vector(X, method='single', metric=[...])
gives the same result as
hclust(dist(X, metric=[...]), method='single')
but uses less memory and is equally fast.
For the Euclidean methods, care must be taken since
hclust expects squared Euclidean
distances. Hence, the call
hclust.vector(X, method='centroid')
is, aside from the lesser memory requirements, equivalent to
d = dist(X) hc = hclust(d^2, method='centroid') hc$height = sqrt(hc$height)
The same applies to the "median" method. The "ward" method in
hclust.vector is equivalent to hclust with method "ward.D2",
but to method "ward.D" only after squaring as above.
More details are in the User's manual
fastcluster.pdf, which is available as
a vignette. Get this from the R command line with
vignette('fastcluster').
Daniel Müllner
https://danifold.net/fastcluster.html
# Taken and modified from stats::hclust ## Perform centroid clustering with squared Euclidean distances, ## cut the tree into ten clusters and reconstruct the upper part of the ## tree from the cluster centers. hc <- hclust.vector(USArrests, "cen") # squared Euclidean distances hc$height <- hc$height^2 memb <- cutree(hc, k = 10) cent <- NULL for(k in 1:10){ cent <- rbind(cent, colMeans(USArrests[memb == k, , drop = FALSE])) } hc1 <- hclust.vector(cent, method = "cen", members = table(memb)) # squared Euclidean distances hc1$height <- hc1$height^2 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)# Taken and modified from stats::hclust ## Perform centroid clustering with squared Euclidean distances, ## cut the tree into ten clusters and reconstruct the upper part of the ## tree from the cluster centers. hc <- hclust.vector(USArrests, "cen") # squared Euclidean distances hc$height <- hc$height^2 memb <- cutree(hc, k = 10) cent <- NULL for(k in 1:10){ cent <- rbind(cent, colMeans(USArrests[memb == k, , drop = FALSE])) } hc1 <- hclust.vector(cent, method = "cen", members = table(memb)) # squared Euclidean distances hc1$height <- hc1$height^2 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)