kPCA reproducibility

[topepo]

I'm not able to reproduce the kernel PCA results in comparison to the underlying function. Here's an example:

library(kernlab)
library(dimRed)

set.seed(131)
tr_dat <- matrix(rnorm(100*6), ncol = 6)
te_dat <- matrix(rnorm(20*6), ncol = 6)
colnames(tr_dat) <- paste0("X", 1:6)
colnames(te_dat) <- paste0("X", 1:6)

k_name <- "rbfdot"
k_par <- list(sigma = .2)

## test values

kpca_obj <- kPCA(stdpars = list(ndim = 3, kernel = k_name, kpar = k_par))
kpca_obj <- kpca_obj@fun(dimRedData(tr_dat), kpca_obj@stdpars)
kpca_pred <- kpca_obj@apply(te_dat)@data

## expected values

kpca_obj_exp <- kpca(tr_dat, 
                     kernel = k_name,
                     kpar = k_par)
kpca_pred_exp <- predict(kpca_obj_exp, tr_dat)[, 1:3]
colnames(kpca_pred_exp) <- paste0("kPCA", 1:3)

I get

> head(kpca_pred)
          kPCA1      kPCA2       kPCA3
[1,] -0.1754955 -2.8205993  0.51416167
[2,]  1.1112348  1.7925091 -0.02363246
[3,]  1.9973353 -0.9198911  0.14218226
[4,]  3.0105551  1.4249128 -2.79424169
[5,] -3.2053340 -2.0046749 -0.79662181
[6,]  1.5522026  3.6696689 -2.54760691
> head(kpca_pred_exp)
         kPCA1      kPCA2      kPCA3
[1,]  2.614505  2.9551241  2.1230302
[2,] -1.827209  2.4680460 -2.3203690
[3,]  2.956935 -1.2295952 -2.9909752
[4,] -3.740879 -0.8210545 -4.0988922
[5,] -1.015746 -1.7453619 -0.5225218
[6,] -2.357748  2.1721046 -2.2195350
> sessionInfo()
R version 3.3.2 (2016-10-31)
Platform: x86_64-apple-darwin13.4.0 (64-bit)
Running under: macOS Sierra 10.12.3

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] dimRed_0.0.3.9001 DRR_0.0.2         CVST_0.2-1        Matrix_1.2-7.1    kernlab_0.9-25   

loaded via a namespace (and not attached):
[1] tools_3.3.2     grid_3.3.2      lattice_0.20-34

BTW what's the best way to access the objects generated in the fun code from the base object? I'd like to get ahold of the PCA rotation matrix or the kPCA object res. That gets computed once on the first call?

Thanks,

Max

[gdkrmr]

you are using tr_dat instead of te_dat in the second line from below, try:

kpca_pred_exp <- predict(kpca_obj_exp, te_dat)[, 1:3]

To get the actual rotation matrix of a pca, you need to look in the environment of its apply function:

emb <- embed(loadDataSet("Iris"), "PCA")
environment(emb@apply)$rot

It even has nice column and row names, for a kPCA object you would have use environment(kpca_obj@apply)$res.

I will make some convenience functions for extracting interesting information from objects.

[gdkrmr]

@topepo maybe you can help with some proposals in #4

[topepo]

My apologies for that; I hate is when I get issues with bad code =[

[gdkrmr]

don't worry, happens to me all the time.