R/d.matrix.R

In thiyangt/fformpp: Feature-based FORecast Model Performance Prediction

#' Make a design matrix for multivariate thin plate splines and others.
#'
#' The number of knots should be smaller than the number of observations.
#'
#' @name d.matrix
#' @title Design matrix for splines
#'
#' @param x "matrix".
#'         the data to be used in the splines. You should *NOT* include a
#'         constant column. The intercept will be added automatically if required in
#'         "args".
#' @param knots "list".
#'         knots locations for different spline components:
#'         knots$thinplate.a: "matrix" Knots for the additive part.
#'         knots$thinplate.s: "matrix" Knots for the surface part.
#' @param splineArgs "list".
#'         Arguments list pass to the function, see bellow.
#'         splineArgs$comp: "character". The combination of splines, which can be
#'         "intercept": If included, will return design matrix with intercept.
#'         "covariates": If included, covariates should be included
#'         "thinplate.s": If included, the thinplate surface will be in;
#'         "thinplate.a": If include, the additive thinplate will be in.
#'         e.g. c("intercept", "covariates", "thinplate.s", "thinplate.a") means all the
#'         four components will be in the design matrix.
#'         splineArgs$thinplate.a.locate: "numeric", c(2, 3, 0) means 2 knots
#'         for first regressor, 3 knots for the second regressor and no knots for third
#'         regressor.
#'
#' @return "matrix".
#'         The dimension depends on the arguments, and the first column is 1 as an intercept
#'         if required in "splineArgs".
#'
#' @references Denison et al (2002)
#' @author Feng Li, Dept. of Statistics, Stockholm University, Sweden.
#' @note First version: Sun Mar  7 13:41:32  CET 2010.
#'       Current:       Tue Dec 28 13:34:39 CET 2010.
#'
#' @export
d.matrix <- function(x,knots,splineArgs)
{

  ## Pre check
  dim.x <- dim(x)
  n <- dim.x[1] # no. of obs.
  m <- dim.x[2] # no. of covs.
  comp <- tolower(splineArgs$comp) # The components of the design matrix
  out.tmp <- list()

  ## Construct the intercept.
  if ("intercept" %in% comp)
  {
    out.tmp[["intercept"]] <- matrix(1, dim.x[1], 1)
  }

  ## Construct the covariates.
  if ("covariates" %in% comp)
  {
    out.tmp[["covariates"]] <- x
  }

  ## Construct the thinplate sufrace
  if ("thinplate.s" %in% comp)
  {
    log.rowDist <- rdist(x, knots[["thinplate.s"]], log = TRUE)
    thinplate.s <- exp(2*log.rowDist)*log.rowDist # P. 141, Denison et al. (2002)

    ## when x -> 0,  x^2log(x)->0 It may produce NaN numerically at this situation.
    thinplate.s[is.nan(thinplate.s)] <- 0
    out.tmp[["thinplate.s"]] <- thinplate.s
  }

  ## Construct additive thinplate spline
  if ("thinplate.a" %in% comp)
  {
    idx <- rep(1:m, splineArgs$thinplate.a.locate)

    x.add <- x[, idx, drop = FALSE]
    knots.add <- matrix(knots[["thinplate.a"]], n, length(idx), byrow = TRUE)

    dist.add <- x.add-knots.add
    thinplate.a <- (dist.add)^2*log(abs(dist.add))

    ## Numerical corrections. x^2*log(x) -> 0 when x -> 0.
    ## But it was NaN numerically.
    thinplate.a[is.nan(thinplate.a)] <- 0

    out.tmp[["thinplate.a"]] <- thinplate.a
  }

  ## Order and combine the output matrix.
  X.out <- matrix(unlist(out.tmp[comp]), n)

  return(X.out)
}