Clusplot - mapping variables

Question

I would like to add to the clusplot graph the variables used for pca as arrows. I am not sure if the method was implemented (I cannot find anything in the documentation).

I created clusplot that looks like this:

With the princomp package princomp I can independently build observations in a similar presentation space with variables (columns) in the form of arrows:

Is there a way to do two things at the same time, showing clusters and pca variables in the same diagram?

Answer 1

I wanted to do the same thing as today, and ended up making fragments from clusplot and biplot . This is a result that can be useful if you want to do the same:

 clusplot2 <- function(dat, clustering, ...) { clusplot(dat, clustering, ...) ## this is from clusplot.default pca <- princomp(dat, scores = TRUE, cor = (ncol(dat) != 2)) ## this is (adapted) from biplot.princomp directions <- t(t(pca$loadings[, 1:2]) * pca$sdev[1:2]) * sqrt(pca$n.obs) ## all below is (adapted) from biplot.default unsigned.range <- function(x) c(-abs(min(x, na.rm = TRUE)), abs(max(x, na.rm = TRUE))) x <- predict(pca)[, 1:2] y <- directions rangx1 <- unsigned.range(x[, 1L]) rangx2 <- unsigned.range(x[, 2L]) rangy1 <- unsigned.range(y[, 1L]) rangy2 <- unsigned.range(y[, 2L]) xlim <- ylim <- rangx1 <- rangx2 <- range(rangx1, rangx2) ratio <- max(rangy1/rangx1, rangy2/rangx2) par(new = T) col <- par("col") if (!is.numeric(col)) col <- match(col, palette(), nomatch = 1L) col <- c(col, col + 1L) cex <- rep(par("cex"), 2) plot(y, axes = FALSE, type = "n", xlim = xlim * ratio, ylim = ylim * ratio, xlab = "", ylab = "", col = col[1L]) axis(3, col = col[2L]) axis(4, col = col[2L]) box(col = col[1L]) text(y, labels = names(dat), cex = cex[2L], col = col[2L]) arrows(0, 0, y[, 1L] * 0.8, y[, 2L] * 0.8, col = col[2L], length = 0.1) } ############################################################ library(cluster) dat <- iris[, 1:4] clus <- pam(dat, k = 3) clusplot2(dat, clus$clustering, main = "Test") 

Of course, there are many opportunities for improvement (as it is simply copied together), but I think that anyone can easily adapt it if necessary.

If you're wondering why arrows (loads * sdev) scale with 0.8 * sqrt (n): I have no idea. I would build * sdev loads, which should resemble the correlation between the main components and variables, but this is how biplot does biplot .

In any case, this should lead to the same arrows as biplot.princomp , and use the same pca for me as clusplot .