CODEWITHSUNDEEP
rdendrogramdendextend
Excelsior
Excelsior

Reputation: 181

strange behaviour with rect.dendrogram in dendextend package

I try to show the influence of the cutoff height on the number of clusters using the iris dataset and visualize the resulting clusters with rect.dendrogram.

if (!require("dendextend")) {install.packages("dendextend")} else {library("dendextend")}

data("iris", package = "datasets")

Data <- list()

Data$Lab <- as.character(iris[,5])
Data$dat <- prcomp(iris[,-5])$x[,1:2]

Data$dist <- dist(Data$dat, method = "euclidean")
Data$hist <- hclust(Data$dist, method = "complete")

# plot dendrogram

hcd <- as.dendrogram(Data$hist)

cluster.height <- 6

par(pty = "m",
    mar = c(1,2,1.5,1),
    mgp = c(1,0,0),
    tck = 0.01,
    cex.axis = 0.75,
    font.main = 1)
plot(sort(hcd), 
     ylab = "Height",
     leaflab = "none")
rect.dendrogram(sort(hcd),
                h = cluster.height,
                border = "black",
                xpd = NA,
                lower_rect = -0.1,
                upper_rect = 0)
abline(h = cluster.height,
       lty = 3)
dev.off()

When using high height values, two rectangles appear.

dendrogram_1

The function searches for the clusters created by the cutoff. The question is whether there is a way to obtain only the larger rectangle? Is there a parameter/option that I have overlooked, or is it a bug in the rect.dendrogram function?

Upvotes: 2

Views: 88

Answers (2)

jay.sf
jay.sf

Reputation: 72583

Looks indeed like some sort of bug to me, at least the function only seems to work with "hclust" objects but not with "dendrogram"s. You could use the relevant parts of the function.

> rect_dnd <- \(tree, which, h, ybadj=0, ytadj=0, ...) {
+   cl <- cutree(tree, h=h)
+   clt <- table(cl)[unique(cl[tree$order])]
+   m <- c(0, cumsum(clt))
+   k <- min(which(rev(tree$height) < h))
+   rect(xleft=m[which] + 0.66,
+        ybottom=mean(rev(tree$height)[(k - 1):k]) + ytadj,
+        xright=m[which + 1] + 0.33,
+        ytop=par()$usr[3] + ybadj, ...)
+ }
> 
> cluster.height <- 6
> 
> par(pty="m", mar=c(1, 2, 1.5, 1), mgp=c(1, 0, 0), tck=0.01, cex.axis=0.75, 
+     font.main=1)
> plot(hcd, ylab="Height", leaflab="none")
> rect_dnd(Data$hist, which=1, h=cluster.height, ytadj=-.1, border='red')
> abline(h=cluster.height, lty=3)

enter image description here

Data:

> Data <- list(Lab=as.character(iris[, 5]), dat=prcomp(iris[, -5])$x[, 1:2])
> Data$dist <- dist(Data$dat, method="euclidean")
> Data$hist <- hclust(Data$dist, method="complete")
> hcd <- as.dendrogram(Data$hist)

Upvotes: 2

I_O
I_O

Reputation: 6911

You could fall back to rect.hclust on which rect.dendrogram is based. rect.hclust allows to also specify which cluster should be highlighted. Example:

cluster.height = 6
pc <- prcomp(iris[-5])

Data <- list(labs = iris$Species,
             data = predict(pc)[,1:2],
             clusters = dat |>
               dist(method = 'euclidean') |>
               hclust(method = 'complete') |>
               sort()
             )

plot(Data$clusters)
rect.hclust(Data$clusters, h = cluster.height, which = 1)

see also ?identify.hclust for interactive identification of dendrogram branches (=clusters)

marking individual cluster

Upvotes: 1

Related Questions