drr {DRR} R Documentation

Dimensionality Reduction via Regression

Description

drr Implements Dimensionality Reduction via Regression using Kernel Ridge Regression.

Usage

drr(
X,
ndim = ncol(X),
lambda = c(0, 10^(-3:2)),
kernel = "rbfdot",
kernel.pars = list(sigma = 10^(-3:4)),
pca = TRUE,
pca.center = TRUE,
pca.scale = FALSE,
fastcv = FALSE,
cv.folds = 5,
fastcv.test = NULL,
fastkrr.nblocks = 4,
verbose = TRUE
)


Arguments

 X input data, a matrix. ndim the number of output dimensions and regression functions to be estimated, see details for inversion. lambda the penalty term for the Kernel Ridge Regression. kernel a kernel function or string, see kernel-class for details. kernel.pars a list with parameters for the kernel. each parameter can be a vector, crossvalidation will choose the best combination. pca logical, do a preprocessing using pca. pca.center logical, center data before applying pca. pca.scale logical, scale data before applying pca. fastcv if TRUE uses fastCV, if FALSE uses CV for crossvalidation. cv.folds if using normal crossvalidation, the number of folds to be used. fastcv.test an optional separate test data set to be used for fastCV, handed over as option test to fastCV. fastkrr.nblocks the number of blocks used for fast KRR, higher numbers are faster to compute but may introduce numerical inaccurracies, see constructFastKRRLearner for details. verbose logical, should the crossvalidation report back.

Details

Parameter combination will be formed and cross-validation used to select the best combination. Cross-validation uses CV or fastCV.

Pre-treatment of the data using a PCA and scaling is made \alpha = Vx. the representation in reduced dimensions is

y_i = \alpha - f_i(\alpha_1, \ldots, \alpha_{i-1})

then the final DRR representation is:

r = (\alpha_1, y_2, y_3, \ldots,y_d)

DRR is invertible by

\alpha_i = y_i + f_i(\alpha_1,\alpha_2, \ldots, alpha_{i-1})

If less dimensions are estimated, there will be less inverse functions and calculating the inverse will be inaccurate.

Value

A list the following items:

• "fitted.data" The data in reduced dimensions.

• "pca.means" The means used to center the original data.

• "pca.scale" The standard deviations used to scale the original data.

• "pca.rotation" The rotation matrix of the PCA.

• "models" A list of models used to estimate each dimension.

• "apply" A function to fit new data to the estimated model.

• "inverse" A function to untransform data.

References

Laparra, V., Malo, J., Camps-Valls, G., 2015. Dimensionality Reduction via Regression in Hyperspectral Imagery. IEEE Journal of Selected Topics in Signal Processing 9, 1026-1036. doi:10.1109/JSTSP.2015.2417833

Examples

tt <- seq(0,4*pi, length.out = 200)
helix <- cbind(
x = 3 * cos(tt) + rnorm(length(tt), sd = seq(0.1, 1.4, length.out = length(tt))),
y = 3 * sin(tt) + rnorm(length(tt), sd = seq(0.1, 1.4, length.out = length(tt))),
z = 2 * tt      + rnorm(length(tt), sd = seq(0.1, 1.4, length.out = length(tt)))
)
helix <- helix[sample(nrow(helix)),] # shuffling data is important!!
system.time(
drr.fit  <- drr(helix, ndim = 3, cv.folds = 4,
lambda = 10^(-2:1),
kernel.pars = list(sigma = 10^(0:3)),
fastkrr.nblocks = 2, verbose = TRUE,
fastcv = FALSE)
)

## Not run:
library(rgl)
plot3d(helix)
points3d(drr.fit$inverse(drr.fit$fitted.data[,1,drop = FALSE]), col = 'blue')
points3d(drr.fit$inverse(drr.fit$fitted.data[,1:2]),             col = 'red')

plot3d(drr.fit$fitted.data) pad <- -3 fd <- drr.fit$fitted.data
xx <- seq(min(fd[,1]),       max(fd[,1]),       length.out = 25)
rgl.linestrips(drr.fit$inverse(cbind(x, y, zz)), col = 'blue') for(y in yy) for(z in zz) rgl.linestrips(drr.fit$inverse(cbind(xx, y, z)), col = 'blue')