prediction.strength in Package fpc - r

Question

I am using the function prediction.strength in the r Package fpc with k-medoids algorithms.
here is my code
prediction.strength(data,2,6,M=10,clustermethod=pamkCBI,DIST,krange=2:6,diss=TRUE,usepam=TRUE)
somehow I get the error message
Error in switch(method, kmeans = kmeans(xdata[indvec[[l]][[i]], ], k, :
EXPR must be a length 1 vector
Does anybody have experience with this r command? There are simple examples like
iriss <- iris[sample(150,20),-5]
prediction.strength(iriss,2,3,M=3,method="pam")
but my problem is that I am using dissimilarity matrix instead of the data itself for the k-medoids algorithms. I don't know how should I correct my code in this case.

Answer 1

Please note that in the package help the following is stated for the prediction.strength:
xdats - data (something that can be coerced into a matrix). Note that this can currently
not be a dissimilarity matrix.
I'm afraid you'll have to hack the function to get it to handle a distance matrix. I'm using the following:
pred <- function (distance, Gmin = 2, Gmax = 10, M = 50,
classification = "centroid", cutoff = 0.8, nnk = 1, ...)
{
require(cluster)
require(class)
xdata <- as.matrix(distance)
n <- nrow(xdata)
nf <- c(floor(n/2), n - floor(n/2))
indvec <- clcenters <- clusterings <- jclusterings <- classifications <- list()
prederr <- list()
dist <- as.matrix(distance)
for (k in Gmin:Gmax) {
prederr[[k]] <- numeric(0)
for (l in 1:M) {
nperm <- sample(n, n)
indvec[[l]] <- list()
indvec[[l]][[1]] <- nperm[1:nf[1]]
indvec[[l]][[2]] <- nperm[(nf[1] + 1):n]
for (i in 1:2) {
clusterings[[i]] <- as.vector(pam(as.dist(dist[indvec[[l]][[i]],indvec[[l]][[i]]]), k, diss=TRUE))
jclusterings[[i]] <- rep(-1, n)
jclusterings[[i]][indvec[[l]][[i]]] <- clusterings[[i]]$clustering
centroids <- clusterings[[i]]$medoids
j <- 3 - i
classifications[[j]] <- classifdist(as.dist(dist), jclusterings[[i]],
method = classification, centroids = centroids,
nnk = nnk)[indvec[[l]][[j]]]
}
ps <- matrix(0, nrow = 2, ncol = k)
for (i in 1:2) {
for (kk in 1:k) {
nik <- sum(clusterings[[i]]$clustering == kk)
if (nik > 1) {
for (j1 in (1:(nf[i] - 1))[clusterings[[i]]$clustering[1:(nf[i] -
1)] == kk]) {
for (j2 in (j1 + 1):nf[i]) if (clusterings[[i]]$clustering[j2] ==
kk)
ps[i, kk] <- ps[i, kk] + (classifications[[i]][j1] ==
classifications[[i]][j2])
}
ps[i, kk] <- 2 * ps[i, kk]/(nik * (nik -
1))
}
}
}
prederr[[k]][l] <- mean(c(min(ps[1, ]), min(ps[2,
])))
}
}
mean.pred <- numeric(0)
if (Gmin > 1)
mean.pred <- c(1)
if (Gmin > 2)
mean.pred <- c(mean.pred, rep(NA, Gmin - 2))
for (k in Gmin:Gmax) mean.pred <- c(mean.pred, mean(prederr[[k]]))
optimalk <- max(which(mean.pred > cutoff))
out <- list(predcorr = prederr, mean.pred = mean.pred, optimalk = optimalk,
cutoff = cutoff, method = clusterings[[1]]$clustermethod,
Gmax = Gmax, M = M)
class(out) <- "predstr"
out
}