This package demonstrates the R package bootSVD (Fisher et al., 2014), using the "Stirling" face dataset, available here:
http://pics.psych.stir.ac.uk/2D_face_sets.htm
The package is currently only meant to be used in demonstrations during talks. The only function available is ex_faces. More customizable examples can be built off of the bootSVD package directly.
To install and run:
## if needed
install.packages("devtools")
## main package
library(devtools)
install_github('aaronjfisher/bootSVD')
install_github('aaronjfisher/ajfisher')
install_github('aaronjfisher/2014_JSM_bootSVD_demo')
## run bootstrap PCA on face dataset
library(bootSVD2014JSMdemo) #note, repo name and package name are not the same
ex_faces(K=3,showFaceEx=TRUE)With standard methods these bootstrap standard errors would require at least 15.7 minutes (see code below). With bootSVD however, exact results are available in approximately 15 seconds.
Approximate timing for the standard bootstrap PCA procedure:
library(bootSVD)
library(bootSvdDemoJSM2014)
# center data matrix before doing PCA
Y<-scale(t(faces_mat),scale=FALSE,center=TRUE)
# Run standard PCA 10 times, then multiply computation time by 100
# to get computation time for 1000 iterations
time_10<-system.time({ #time it 10 times
for(i in 1:10)
dump<-fastSVD(Y,nv=3)
})
# minutes required for 1000 traditional bootstrap
time_10['elapsed'] * 100 / 60
# Approx 15.7 minReferences:
Aaron Fisher, Brian Caffo, and Vadim Zipunnikov. Fast, Exact Bootstrap Principal Component Analysis for p>1 million. Working Paper, 2014. http://arxiv.org/abs/1405.0922