ANTsR (version 0.3.1)

sparseDecom: Convenience wrapper for eigenanatomy decomposition.

Description

Decomposes a matrix into sparse eigenevectors to maximize explained variance.

Usage

sparseDecom(inmatrix = NA, inmask = 0, sparseness = 0.01, nvecs = 50,
  its = 5, cthresh = 250, statdir = NA, z = 0, smooth = 0,
  initializationList = list(), mycoption = 0, robust = 0, ell1 = 1,
  getSmall = 0)

Arguments

inmatrix

n by p input images , subjects or time points by row , spatial variable lies along columns

inmask

optional antsImage mask

sparseness

lower values equal more sparse

nvecs

number of vectors

its

number of iterations

cthresh

cluster threshold

statdir

place on disk to save results

z

u penalty, experimental

smooth

smoothness eg 0.5

initializationList

see initializeEigenanatomy

mycoption

0, 1 or 2 all produce different output 0 is combination of 1 (spatial orthogonality) and 2 (subject space orthogonality)

robust

rank transform input data - good for data checking

ell1

the ell1 grad descent param

getSmall

try to get smallest evecs (bool)

Value

outputs a decomposition of a population or time series matrix

Examples

# NOT RUN {
mat<-replicate(100, rnorm(20))
mydecom<-sparseDecom( mat )
# }
# NOT RUN {
# for prediction
if ( usePkg("randomForest") & usePkg("spls")  & usePkg('BGLR') ) {
data(lymphoma) # from spls
training<-sample( rep(c(TRUE,FALSE),31)  )
sp<-0.02 ; myz<-0
ldd<-sparseDecom( lymphoma$x[training,], nvecs=5 , sparseness=( sp ),
  mycoption=1, z=myz ) # NMF style
traindf<-data.frame( lclass=as.factor(lymphoma$y[ training  ]),
  eig = lymphoma$x[training,]  %*% as.matrix(ldd$eigenanatomyimages ))
testdf<-data.frame(  lclass=as.factor(lymphoma$y[ !training ]),
 eig = lymphoma$x[!training,] %*% as.matrix(ldd$eigenanatomyimages ))
myrf<-randomForest( lclass ~ . ,   data=traindf )
predlymp<-predict(myrf, newdata=testdf)
print(paste('N-errors:',sum(abs( testdf$lclass != predlymp ) ),
  ' non-zero ',sum(abs( ldd$eigenanatomyimages ) > 0 ) ) )
# compare to http://arxiv.org/pdf/0707.0701v2.pdf
# now SNPs
data(mice)
snps<-quantifySNPs( mice.X, shiftit = TRUE )
numericalpheno<-as.matrix( mice.pheno[,c(4,5,13,15) ] )
nfolds<-6
train<-sample( rep( c(1:nfolds), 1800/nfolds ) )
train<-( train < 4 )
lrmat<-lowrankRowMatrix( as.matrix( snps[train,] ) ,  50 )
snpd<-sparseDecom( lrmat, nvecs=20 , sparseness=( 0.001), z=-1 )
projmat<-as.matrix( snpd$eig )
snpse<-as.matrix( snps[train, ]  ) %*% projmat
traindf<-data.frame( bmi=numericalpheno[train,3] , snpse=snpse)
snpse<-as.matrix( snps[!train, ]  ) %*% projmat
testdf <-data.frame( bmi=numericalpheno[!train,3] , snpse=snpse )
myrf<-randomForest( bmi ~ . , data=traindf )
preddf<-predict(myrf, newdata=testdf )
cor.test(preddf, testdf$bmi )
plot(preddf, testdf$bmi )
}
# }