R/adjusttensor.R

In TULIP: A Toolbox for Linear Discriminant Analysis with Penalties

adjten <- function(x, z, y, testx=NULL, testz=NULL, is.centered=FALSE){
  

  n <- length(x)
  dimen <- dim(x[[1]])
  nvars <- prod(dimen)
  vec_x <- matrix(0, nvars, n)
  for (i in 1:n){
    vec_x[,i] <- matrix(x[[i]], nvars, 1)
  }
  vec_x <- t(vec_x)
  
  if (!is.null(testx)){
    tesize <- length(testx)
    vec_testx <- matrix(0, nvars, tesize)
    for (i in 1:tesize){
      vec_testx[,i] <- matrix(testx[[i]], nvars, 1)
    }
    vec_testx <- t(vec_testx)
  }
  z <- as.matrix(z)
  cz <- z
  n <- dim(z)[1]
  q <- dim(z)[2]
  cvecx <- vec_x
  nclass <- as.integer(length(unique(y)))
  if (is.centered==FALSE){
    for (i in 1:nclass){
      if (q>1) {cz[y == i,] <- sweep(z[y == i,],2,colMeans(z[y == i,]))}
      else {cz[y==i] <- z[y==i]-mean(z[y==i])}
      cvecx[y == i,] <- sweep(vec_x[y == i,],2,colMeans(vec_x[y == i,]))
    }
  }
  c <- solve(t(cz) %*% cz) %*% t(cz)
  c <- c %*% cvecx

  vec_xres <- vec_x-z%*%c

  xres <- array(list(),n)
  for (i in 1:n){
    xres[[i]] <- array(vec_xres[i,],dim=dimen)
  }
 
  
  if (!is.null(testx)){
    vec_testxres <- vec_testx-testz %*% c
    testxres <- array(list(), tesize)
    for (i in 1:tesize){
      testxres[[i]] <- array(vec_testxres[i,], dim=dimen)
    } 
  }else{
    testxres <- NULL
  }
  
  muz <- matrix(0, nrow=q, ncol=(nclass-1))
  for (i in 2:nclass){
    if (q>1){muz[,(i-1)] <- apply(z[y==i,], 2, mean)-apply(z[y==1,], 2, mean)}
    else {muz[i-1] <- mean(z[y==i])-mean(z[y==1])}
  }
  gamma <- solve(cov(z)) %*% muz
  
  outlist = list(gamma=gamma,xres=xres,testxres=testxres)
  outlist
}