R/grnn.x_pfi.R

In yager: Yet Another General Regression Neural Network

#' Derive the permutation feature importance of a predictor used in the GRNN 
#'
#' The function \code{grnn.x_pfi} derives the permutation feature importance (PFI) of a predictor used in the GRNN
#'
#' @param net   The GRNN object generated by grnn.fit() 
#' @param i     The ith predictor in the GRNN
#' @param class TRUE or FALSE, whether it is for the classification or not
#' @param ntry  The number of random permutations to try, 1e3 times by default
#' @param seed  The seed value for the random permutation
#'
#' @return A vector with the variable name and the PFI value.
#'
#' @seealso \code{\link{grnn.x_imp}}
#'
#' @examples
#' data(iris, package = "datasets")
#' Y <- ifelse(iris[, 5] == "setosa", 1, 0)
#' X <- scale(iris[, 1:4])
#' gnet <- grnn.fit(x = X, y = Y)
#' grnn.x_pfi(net = gnet, 1)

grnn.x_pfi <- function(net, i, class = FALSE, ntry = 1e3, seed = 1) {
  if (class(net) != "General Regression Neural Net") stop("net needs to be a GRNN.", call. = F)
  if (!(class %in% c(TRUE, FALSE))) stop("the class input is not correct.", call. = F)

  xname <- colnames(net$x)[i]
  set.seed(seed)
  seeds <- floor(runif(ntry) * 1e8)
  ol <- lapply(seeds, function(s) with(set.seed(s), sample(seq(nrow(net$x)), nrow(net$x), replace = F)))
  cl <- Reduce(c, lapply(ol, function(o) abs(cor(seq(nrow(net$x)), o))))
  x <- net$x
  x[, i] <-  net$x[ol[[which(cl == min(cl))]], i]
  if (class == TRUE) {
    auc0 <- MLmetrics::AUC(grnn.predict(net, net$x), net$y)
    auc1 <- MLmetrics::AUC(grnn.predict(net, x), net$y)
    pfi  <- round(max(0, 1 - auc1 / auc0), 8)
  } else {
    rsq0 <- MLmetrics::R2_Score(grnn.predict(net, net$x), net$y)
    rsq1 <- MLmetrics::R2_Score(grnn.predict(net, x), net$y)
    pfi  <- round(max(0, 1 - rsq1 / rsq0), 8)
  }
  return(data.frame(var = xname, pfi = pfi))
}