5
$\begingroup$

I would like to remove the first principal component from a data set, but keep that data set in its original coordinates. I have taken a stab at this by taking PCA, zeroing the first PC, and then rotating back using the inverse of the eigvenvector matrix. Is that the most efficient way to do this?

create a sample data set:

set.seed(1234)
xx <- rnorm(500)
yy <- 0.5 * xx + rnorm(500, sd = 0.3)
vec <- cbind(xx, yy)
plot(vec, xlim = c(-4, 4), ylim = c(-4, 4))

enter image description here

Take principal components and zero out the first PC:

vv <- eigen(cov(vec))$vectors
newvec <- vec %*% vv
newvec[, 1] <- 0

Now rotate the new data set back to its original coordinates using the inverse of the PCA rotation matrix:

rvec <- newvec %*% t(vv) # transpose of orthogonal matrix = inverse
# plot new points in red and plot eigenvectors in green
points(rvec, col = "red")
arrows(0, 0, vv[1,1], vv[2, 1], col = "green3", lwd = 2)
arrows(0, 0, vv[1,2], vv[2, 2], col = "green3", lwd = 2)
legend("topleft", legend = c("original data", "data after extracting PC1", "eigenvectors of original data"), fill = c("black", "red", "green3"))

enter image description here

As you can see it seems to agree with the eigenvector orientations in green. But is this the correct and/or best way? Can I avoid the intermediate matrix multiplications for example?

| cite | improve this question | |
$\endgroup$
  • $\begingroup$ +1 Well formulated and illustrated question. Your vv is an orthogonal matrix, hence its inverse equals its transpose. No need to compute the inverse. Answer: correct but not the best :) $\endgroup$ – amoeba Aug 4 '16 at 11:33
  • $\begingroup$ duly edited to replace solve() with t() $\endgroup$ – Thomas Browne Aug 4 '16 at 11:47
  • 2
    $\begingroup$ Cool, but now there is nothing to answer anymore. The answer is Yes. $\endgroup$ – amoeba Aug 4 '16 at 11:50
  • $\begingroup$ Just out of curiosity: after determining vv, shouldn't this also be possible in a single step instead of transforming all data to PC space and back? This would boil down to projecting data in the original space onto the subspace represented by the red line in the figure above, which is known after computing vv, and therefore shouldn't be too tricky, right? $\endgroup$ – geekoverdose Aug 4 '16 at 12:43
  • $\begingroup$ @geekoverdose Yes, but this "single step" will involve projecting $X$ via $WW^\top$, i.e. computing $XWW^\top$, which is not that different from computing $XW$ and then multiplying with $W^\top$ (here $W$ denotes vv, or rather all columns of vv apart from the first). $\endgroup$ – amoeba Aug 4 '16 at 13:02

Browse other questions tagged or ask your own question.