Search results in numerous places report that the argument nstart in R's function kmeans sets a number of iterations of the algorithm and chooses 'the best one', see e.g. https://datascience.stackexchange.com/questions/11485/k-means-in-r-usage-of-nstart-parameter. Can anyone provide any clarity on how it does this, i.e. by what measure does it define best?
Secondly: R's kmeans function takes an argument centers. Here, as typical in k-means, it is possible to initialise the centroids before the algorithm begins expectation-maximisation, by choosing as initial centroids rows (data-points) from within your data-set. (You could supply, in vector form, points not present in your data-set as well, with considerably greater effort. In this case you could in theory choose the global optimum as your centroids. This is not what I'm asking for.) When nstart or the seed randomises initializations, I am quite sure that it does so by picking a random choice of centroids from your data-set and starting from those (not just a random set of points within the space).
In general, therefore, I'm looking for a way to get a good (e.g. best out of $n$ trials, or best from nstart) set of starting data-instances from the data-set as initial centroids. Is there any way of extracting the 'winning' (=best) set of initial centroids from nstart (which I could then use, say, in the centers parameter in future)? Any other streamlined & quick way to get a very good set of starting centroids (presumably, reasonably close to where the cluster centres will end up being)?
Is there perhaps, at least, a way to extract from a given kmeans run, what initial centroids it chose to start with?
The criterion that kmeans tries to minimize is the trace of the within scatter matrix, i.e. (unfortunately, this forum does not support LaTeX, but you hopefully can read it nevertheless):
$$ trace(S_w) = \sum_{k=1}^K \sum{x \in C_k} ||x - \mu_k||^2 $$
Concerning the best starting point: obviously, the "best" starting point would be the cluster centers eventually chosen by kmeans. These are returned in the attribute centers:
km <- kmeans(iris[,-5], 3)
print(km$centers)
If you are looking for the best random start point, you can create random start points yourself (with runif), do this nstart times and evaluate which initial configuration leads to the smallest km$tot.withinss:
nstart <- 10
K <- 3 # number of clusters
D <- 4 # data point dimension
# select possible range
r.min <- apply(iris[,-5], MARGIN=2, FUN=min)
r.max <- apply(iris[,-5], MARGIN=2, FUN=max)
for (i in 1:nstart) {
centers <- data.frame(runif(K, r.min[d], r.max[d]))
for (d in 2:D) {
centers <- cbind(centers, data.frame(runif(K, r.min[d], r.max[d])))
}
names(centers) <- names(iris[,-5])
# call kmeans with centers and compare tot.withinss
# ...
}
I am trying to cluster a Multidimensional Functional Object with the "kmeans" algorithms. What does it mean: So I don't have anymore a vector per each row or Individual, even more a 3x3 observation matrix per each Individual.For example: Individual = 1 has the following observations:
(x1, x2, x3),(y1,y2,y3),(z1,z2,z3).
The same structure of observations is also given for the other Individuals. So do you know how I can cluster with "kmeans" including all 3 observation vectors -and not only one observation vector how it is normal used for "kmeans" clustering?
Would you do it for each observation vector, f.e. (x1, x2, x3), separately and then combine the Information somehow together? I want to do this with the kmeans() Function in R.
Many thanks for your answers!
Using k-means you interpret each observation as a point in an N-dimensional vector space. Then you minimize the distances between your observations and the cluster centers.
Since, the data is viewed as dots in an N-dim space, the actual arrangement of the values does not matter.
You can, therefore, either tell your k-means routine to use a matrix norm, for example the Frobenius norm, to compute the distances. The other way would be to flatten your observations from 3 by 3 matrices to 1 by 9 vectors. The Frobenius norm of a NxN matrix is equivalent to the euclidean norm of a 1xN^2 vector.
Just give the argument to kmeans() with all the three columns it'll calculate the distances in 3 dimension, if that is what you are looking for.
I am ecologist, using mainly the vegan R package.
I have 2 matrices (sample x abundances) (See data below):
matrix 1/ nrow= 6replicates*24sites, ncol=15 species abundances (fish)
matrix 2/ nrow= 3replicates*24sites, ncol=10 species abundances (invertebrates)
The sites are the same in both matrices. I want to get the overall bray-curtis dissimilarity (considering both matrices) among pairs of sites. I see 2 options:
option 1, averaging over replicates (at the site scale) fishes and macro-invertebrates abundances, cbind the two mean abundances matrix (nrow=24sites, ncol=15+10 mean abundances) and calculating bray-curtis.
option 2, for each assemblage, computing bray-curtis dissimilarity among pairs of sites, computing distances among sites centroids. Then summing up the 2 distance matrix.
In case I am not clear, I did these 2 operations in the R codes below.
Please, could you tell me if the option 2 is correct and more appropriate than option 1.
thank you in advance.
Pierre
here is below the R code exemples
generating data
library(plyr);library(vegan)
#assemblage 1: 15 fish species, 6 replicates per site
a1.env=data.frame(
Habitat=paste("H",gl(2,12*6),sep=""),
Site=paste("S",gl(24,6),sep=""),
Replicate=rep(paste("R",1:6,sep=""),24))
summary(a1.env)
a1.bio=as.data.frame(replicate(15,rpois(144,sample(1:10,1))))
names(a1.bio)=paste("F",1:15,sep="")
a1.bio[1:72,]=2*a1.bio[1:72,]
#assemblage 2: 10 taxa of macro-invertebrates, 3 replicates per site
a2.env=a1.env[a1.env$Replicate%in%c("R1","R2","R3"),]
summary(a2.env)
a2.bio=as.data.frame(replicate(10,rpois(72,sample(10:100,1))))
names(a2.bio)=paste("I",1:10,sep="")
a2.bio[1:36,]=0.5*a2.bio[1:36,]
#environmental data at the sit scale
env=unique(a1.env[,c("Habitat","Site")])
env=env[order(env$Site),]
OPTION 1, averaging abundances and cbind
a1.bio.mean=ddply(cbind(a1.bio,a1.env),.(Habitat,Site),numcolwise(mean))
a1.bio.mean=a1.bio.mean[order(a1.bio.mean$Site),]
a2.bio.mean=ddply(cbind(a2.bio,a2.env),.(Habitat,Site),numcolwise(mean))
a2.bio.mean=a2.bio.mean[order(a2.bio.mean$Site),]
bio.mean=cbind(a1.bio.mean[,-c(1:2)],a2.bio.mean[,-c(1:2)])
dist.mean=vegdist(sqrt(bio.mean),"bray")
OPTION 2, computing for each assemblage distance among centroids and summing the 2 distances matrix
a1.dist=vegdist(sqrt(a1.bio),"bray")
a1.coord.centroid=betadisper(a1.dist,a1.env$Site)$centroids
a1.dist.centroid=vegdist(a1.coord.centroid,"eucl")
a2.dist=vegdist(sqrt(a2.bio),"bray")
a2.coord.centroid=betadisper(a2.dist,a2.env$Site)$centroids
a2.dist.centroid=vegdist(a2.coord.centroid,"eucl")
summing up the two distance matrices using Gavin Simpson 's fuse()
dist.centroid=fuse(a1.dist.centroid,a2.dist.centroid,weights=c(15/25,10/25))
summing up the two euclidean distance matrices (thanks to Jari Oksanen correction)
dist.centroid=sqrt(a1.dist.centroid^2 + a2.dist.centroid^2)
and the 'coord.centroid' below for further distance-based analysis (is it correct ?)
coord.centroid=cmdscale(dist.centroid,k=23,add=TRUE)
COMPARING OPTION 1 AND 2
pco.mean=cmdscale(vegdist(sqrt(bio.mean),"bray"))
pco.centroid=cmdscale(dist.centroid)
comparison=procrustes(pco.centroid,pco.mean)
protest(pco.centroid,pco.mean)
An easier solution is just to flexibly combine the two dissimilarity matrices, by weighting each matrix. The weights need to sum to 1. For two dissimilarity matrices the fused dissimilarity matrix is
d.fused = (w * d.x) + ((1 - w) * d.y)
where w is a numeric scalar (length 1 vector) weight. If you have no reason to weight one of the sets of dissimilarities more than the other, just use w = 0.5.
I have a function to do this for you in my analogue package; fuse(). The example from ?fuse is
train1 <- data.frame(matrix(abs(runif(100)), ncol = 10))
train2 <- data.frame(matrix(sample(c(0,1), 100, replace = TRUE),
ncol = 10))
rownames(train1) <- rownames(train2) <- LETTERS[1:10]
colnames(train1) <- colnames(train2) <- as.character(1:10)
d1 <- vegdist(train1, method = "bray")
d2 <- vegdist(train2, method = "jaccard")
dd <- fuse(d1, d2, weights = c(0.6, 0.4))
dd
str(dd)
This idea is used in supervised Kohonen networks (supervised SOMs) to bring multiple layers of data into a single analysis.
analogue works closely with vegan so there won't be any issues running the two packages side by side.
The correctness of averaging distances depends on what are you doing with those distances. In some applications you may expect that they really are distances. That is, they satisfy some metric properties and have a defined relation to the original data. Combined dissimilarities may not satisfy these requirements.
This issue is related to the controversy of partial Mantel type analysis of dissimilarities vs. analysis of rectangular data that is really hot (and I mean red hot) in studies of beta diversities. We in vegan provide tools for both, but I think that in most cases analysis of rectangular data is more robust and more powerful. With rectangular data I mean normal sampling units times species matrix. The preferred dissimilarity based methods in vegan map dissimilarities onto rectangular form. These methods in vegan include db-RDA (capscale), permutational MANOVA (adonis) and analysis of within-group dispersion (betadisper). Methods working with disismilarities as such include mantel, anosim, mrpp, meandis.
The mean of dissimilarities or distances usually has no clear correspondence to the original rectangular data. That is: mean of the dissimilarities does not correspond to the mean of the data. I think that in general it is better to average or handle data and then get dissimilarities from transformed data.
If you want to combine dissimilarities, analogue::fuse() style approach is most practical. However, you should understand that fuse() also scales dissimilarity matrices into equal maxima. If you have dissimilarity measures in scale 0..1, this is usually minor issue, unless one of the data set is more homogeneous and has a lower maximum dissimilarity than others. In fuse() they are all equalized so that it is not a simple averaging but averaging after range equalizing. Moreover, you must remember that averaging dissimilarities usually destroys the geometry, and this will matter if you use analysis methods for rectangularized data (adonis, betadisper, capscale in vegan).
Finally about geometry of combining dissimilarities. Dissimilarity indices in scale 0..1 are fractions of type A/B. Two fractions can be added (and then divided to get the average) directly only if the denominators are equal. If you ignore this and directly average the fractions, then the result will not be equal to the same fraction from averaged data. This is what I mean with destroying geometry. Some open-scaled indices are not fractions and may be additive. Manhattan distances are additive. Euclidean distances are square roots of squared differences, and their squares are additive but not the distances directly.
I demonstrate these things by showing the effect of adding together two dissimilarities (and averaging would mean dividing the result by two, or by suitable weights). I take the Barro Colorado Island data of vegan and divide it into two subsets of slightly unequal sizes. A geometry preserving addition of distances of subsets of the data will give the same result as the analysis of the complete data:
library(vegan) ## data and vegdist
library(analogue) ## fuse
data(BCI)
dim(BCI) ## [1] 50 225
x1 <- BCI[, 1:100]
x2 <- BCI[, 101:225]
## Bray-Curtis and fuse: not additive
plot(vegdist(BCI), fuse(vegdist(x1), vegdist(x2), weights = c(100/225, 125/225)))
## summing distances is straigthforward (they are vectors), but preserving
## their attributes and keeping the dissimilarities needs fuse or some trick
## like below where we make dist structure dtmp to be replaced with the result
dtmp <- dist(BCI) ## dist skeleton with attributes
dtmp[] <- dist(x1, "manhattan") + dist(x2, "manhattan")
## manhattans are additive and can be averaged
plot(dist(BCI, "manhattan"), dtmp)
## Fuse rescales dissimilarities and they are no more additive
dfuse <- fuse(dist(x1, "man"), dist(x2, "man"), weights=c(100/225, 125/225))
plot(dist(BCI, "manhattan"), dfuse)
## Euclidean distances are not additive
dtmp[] <- dist(x1) + dist(x2)
plot(dist(BCI), dtmp)
## ... but squared Euclidean distances are additive
dtmp[] <- sqrt(dist(x1)^2 + dist(x2)^2)
plot(dist(BCI), dtmp)
## dfuse would rescale squared Euclidean distances like Manhattan (not shown)
I only considered addition above, but if you cannot add, you cannot average. It is a matter of taste if this is important. Brave people will average things that cannot be averaged, but some people are more timid and want to follow the rules. I rather go the second group.
I like this simplicity of this answer, but it only applies to adding 2 distance matrices:
d.fused = (w * d.x) + ((1 - w) * d.y)
so I wrote my own snippet to combine an array of multiple distance matrices (not just 2), and using standard R packages:
# generate array of distance matrices
x <- matrix(rnorm(100), nrow = 5)
y <- matrix(rnorm(100), nrow = 5)
z <- matrix(rnorm(100), nrow = 5)
dst_array <- list(dist(x),dist(y),dist(z))
# create new distance matrix with first element of array
dst <- dst_array[[1]]
# loop over remaining array elements, add them to distance matrix
for (jj in 2:length(dst_array)){
dst <- dst + dst_array[[jj]]
}
You could also use a vector of similar size to dst_array in order to define scaling factors
dst <- dst + my_scale[[jj]] * dst_array[[jj]]
I need to run clustering on the correlations of data row vectors, that is, instead of using individual variables as clustering predictor variables, I intend to use the correlations between the vector of variables between data rows.
Is there a function in R that does vector-based clustering. If not and I need to do it manually, what is the right data format to feed in a function such as cmeans or kmeans?
Say, I have m variables and n data rows, the m variables constitute one vector for each data row. so I have a n X n matrix for correlation or cosine. Can this matrix be plugged in the clustering function directly or certain processing is required?
Many thanks.
You can transform your correlation matrix into a dissimilarity matrix,
for instance 1-cor(x) (or 2-cor(x) or 1-abs(cor(x))).
# Sample data
n <- 200
k <- 10
x <- matrix( rnorm(n*k), nr=k )
x <- x * row(x) # 10 dimensions, with less information in some of them
# Clustering
library(cluster)
r <- pam(1-cor(x), diss=TRUE, k=5)
# Check the results
plot(prcomp(t(x))$x[,1:2], col=r$clustering, pch=16, cex=3)
R clustering is often a bit limited. This is a design limitation of R, since it heavily relies on low-level C code for performance. The fast kmeans implementation included with R is an example of such a low-level code, that in turn is tied to using Euclidean distance.
There are a dozen of extensions and alternatives available in the community around R. There are PAM, CLARA and CLARANS for example. They aren't exactly k-means, but closely related. There should be a "spherical k-means" somewhere, that is sensible for cosine distance. There is the whole family of hierarchical clusterings (which scale rather badly - usually O(n^3), with O(n^2) in a few exceptions - but are very easy to understand conceptually).
If you want to explore some more clustering options, have a look at ELKI, it should allow clustering (with various methods, including k-means) by correlation based distances (and it also includes such distance functions). It's not R, though, but Java. So if you are bound to using R, it won't work for you.
I have a spatial dataframe with about 3000 points. I want to generate a matrix that provides the k (in this case 30) nearest neighbors for each point.
I can do it using a loop but i feel that there should be an elegant and optimal way for spatial points dataframe class that i do not know of.
Probably the fastest is to use RANN package - assuming you have x and y:
library(RANN)
m <- as.matrix(nn(data.frame(x=x, y=y, z=rep(0,length(x))), p=30)$nn.idx)
gives you a 3000 x 30 matrix of closest neighbors. It is several orders of magnitude faster than a naive quadratic search.
Edit: Just for completeness, it doesn't matter which ANN frontend you pick, with FNN (suggested by Spacedman) this would be
library(FNN)
m <- get.knn(data.frame(x=x, y=y), 30)$nn.index