R/linear_LSIR.R

In kisungyou/Rdimtools: Dimension Reduction and Estimation Methods

#' Localized Sliced Inverse Regression
#'
#' Localized SIR (SIR) is an extension of celebrated SIR method. As its name suggests,
#' the \emph{locality} concept is brought in that for each slice, only local data points
#' are considered in order to discover intrinsic structure of the data.
#'
#' @param X an \eqn{(n\times p)} matrix or data frame whose rows are observations
#' and columns represent independent variables.
#' @param response a length-\eqn{n} vector of response variable.
#' @param ndim an integer-valued target dimension.
#' @param h the number of slices to divide the range of response vector.
#' @param preprocess an additional option for preprocessing the data.
#' Default is "center". See also \code{\link{aux.preprocess}} for more details.
#' @param ycenter a logical; \code{TRUE} to center the response variable, \code{FALSE} otherwise.
#' @param numk size of determining neighborhood via \eqn{k}-nearest neighbor selection.
#' @param tau regularization parameter for adjusting rank-deficient scatter matrix.
#'
#' @return a named list containing
#' \describe{
#' \item{Y}{an \eqn{(n\times ndim)} matrix whose rows are embedded observations.}
#' \item{trfinfo}{a list containing information for out-of-sample prediction.}
#' \item{projection}{a \eqn{(p\times ndim)} whose columns are basis for projection.}
#' }
#'
#' @examples
#' ## generate swiss roll with auxiliary dimensions
#' ## it follows reference example from LSIR paper.
#' set.seed(100)
#' n     = 123
#' theta = runif(n)
#' h     = runif(n)
#' t     = (1+2*theta)*(3*pi/2)
#' X     = array(0,c(n,10))
#' X[,1] = t*cos(t)
#' X[,2] = 21*h
#' X[,3] = t*sin(t)
#' X[,4:10] = matrix(runif(7*n), nrow=n)
#'
#' ## corresponding response vector
#' y = sin(5*pi*theta)+(runif(n)*sqrt(0.1))
#'
#' ## try different number of neighborhoods
#' out1 = do.lsir(X, y, numk=5)
#' out2 = do.lsir(X, y, numk=10)
#' out3 = do.lsir(X, y, numk=25)
#'
#' ## visualize
#' opar <- par(no.readonly=TRUE)
#' par(mfrow=c(1,3))
#' plot(out1$Y, main="LSIR::nbd size=5")
#' plot(out2$Y, main="LSIR::nbd size=10")
#' plot(out3$Y, main="LSIR::nbd size=25")
#' par(opar)
#'
#' @references
#' \insertRef{wu_localized_2010}{Rdimtools}
#'
#' @seealso \code{\link{do.sir}}
#' @author Kisung You
#' @rdname linear_LSIR
#' @concept linear_methods
#' @export
do.lsir <- function(X, response, ndim=2, h=max(2, round(nrow(X)/5)),
                    preprocess=c("center","scale","cscale","decorrelate","whiten"),
                    ycenter=FALSE, numk = max(2, round(nrow(X)/10)), tau=1.0){
  #------------------------------------------------------------------------
  ## PREPROCESSING
  #   1. data matrix
  aux.typecheck(X)
  n = nrow(X)
  p = ncol(X)
  #   2. response
  response = as.double(response)
  if ((any(is.infinite(response)))||(!is.vector(response))||(any(is.na(response)))){
    stop("* do.sir : 'response' should be a vector containing no NA values.")
  }
  #   3. ndim
  ndim = as.integer(ndim)
  if (!check_ndim(ndim,p)){stop("* do.lsir : 'ndim' is a positive integer in [1,#(covariates)).")}
  #   4. h : number of slices
  h = as.integer(h)
  if (!is.factor(response)){
    if (!check_NumMM(h,2,ceiling(n/2),compact=TRUE)){stop("* do.lsir : the number of slices should be in [2,n/2].")}
  }  #   5. preprocess
  if (missing(preprocess)){
    algpreprocess = "center"
  } else {
    algpreprocess = match.arg(preprocess)
  }
  #   6. numk
  numk = as.integer(numk)
  if (!check_NumMM(numk,1,(n-1),compact=TRUE)){stop("* do.lsir : 'numk' should be in [1,n-1].")}
  #   7. tau
  tau = as.double(tau)
  if (!check_NumMM(tau,0,Inf,compact=TRUE)){stop("* do.lsir : 'tau' should be a nonnegative real number.")}


  #------------------------------------------------------------------------
  ## COMPUTATION : PRELIMINARY
  #   1. preprocessing of data
  tmplist = aux.preprocess.hidden(X,type=algpreprocess,algtype="linear")
  trfinfo = tmplist$info
  pX      = tmplist$pX

  if (!is.logical(ycenter)){
    stop("* do.lsir : 'ycenter' should be a logical variable.")
  }
  if (ycenter==TRUE){
    response = response-mean(response)
  }


  #   2. build label matrix
  if (!is.factor(response)){
    label  = as.integer(sir_makelabel(response, h))
  } else {
    label = as.integer(response)
  }
  ulabel = unique(label)
  nlabel = length(ulabel)

  #   3. compute classwise and overall mean
  class_mean  = array(0,c(nlabel,p))
  class_count = rep(0,nlabel)
  for (i in 1:nlabel){
    idxclass = which(label==ulabel[i])
    class_mean[i,] = as.vector(colMeans(pX[idxclass,]) )
    class_count[i] = length(idxclass)
  }
  all_mean = as.vector(colMeans(pX))
  #   4. compute Empirical Covariance
  mat_Sigma = aux_scatter(pX, all_mean)/n
  #   5. compute Empirical Between-ness Covariance with localization
  mean_local = array(0,c(n,p))
  for (i in 1:n){
    # 5-1. target label index
    tgtlabel = setdiff(which(label==label[i]), i)
    # 5-2. get the logical of smallest distances
    tvec = as.vector(pX[i,])
    tmat = pX[tgtlabel,]
    partidx  = lsir_smallest(tvec, tmat, numk)
    partlabel = tgtlabel[partidx]
    # 5-3. compute the mean
    mean_local[i,] = as.vector(colMeans(pX[partlabel,]))
  }
  # 5-4. compute mean scatter
  mat_Gamma = aux_scatter(mean_local, all_mean)/n + tau*diag(p)


  #------------------------------------------------------------------------
  ## COMPUTATION : MAIN COMPUTATION
  #   1. do matrix inversion.. I hate it.
  costInv = aux.bicgstab(mat_Sigma, mat_Gamma, verbose=FALSE)$x
  #   2. find top eigenvectors
  projection = aux.adjprojection(RSpectra::eigs(costInv, ndim)$vectors)


  #------------------------------------------------------------------------
  ## RETURN
  result = list()
  result$Y = pX%*%projection
  result$trfinfo = trfinfo
  result$projection = projection
  return(result)
}




#  ------------------------------------------------------------------------
#' @keywords internal
#' @noRd
lsir_smallest <- function(tvec, tmat, numk){
  n = nrow(tmat)
  distvec = rep(0,n)
  for (i in 1:n){
    vecdiff = tvec-as.vector(tmat[i,])
    distvec[i] = sum(vecdiff*vecdiff)
  }
  orderdist = order(distvec)
  orderpart = orderdist[1:max(min(n,numk),1)]

  logicalvec = rep(FALSE,n)
  logicalvec[orderpart] = TRUE
  return(logicalvec)
}