R/metrics.kappa.max.R

In Laurae2/LauraeDS: Laurae's Data Science Package

#' Maximum Kappa statistic
#' 
#' This function allows to use a custom thresholding method to maximize the Kappa statistic.
#' 
#' @param preds Type: numeric. The predictions.
#' @param labels Type: numeric. The labels (0, 1).
#' @param threshold Type: logical. Whether to return the threshold. Defaults to \code{FALSE}.
#' 
#' @return A one or two element vector containing the maximum Kappa statistic for binary data, and the threshold used if requested. Returns \code{-1} in case of errors.
#' 
#' @export

metrics.kappa.max <- function(preds, labels, threshold = FALSE) {
  
  DT <- data.table(y_true = labels, y_prob = preds, key = "y_prob")
  cleaner <- !duplicated(DT[, "y_prob"], fromLast = TRUE)
  nump <- sum(labels)
  counter <- length(labels)
  numn <- counter - nump
  
  DT[, tn_v := as.numeric(cumsum(y_true == 0))]
  DT[, fp_v := cumsum(y_true == 1)]
  DT[, fn_v := numn - tn_v]
  DT[, tp_v := nump - fp_v]
  DT <- DT[cleaner, ]
  DT <- DT[, pObs := (tp_v + tn_v) / counter]
  DT <- DT[, pExp := (((tp_v + fn_v) * (tp_v + fp_v)) + ((fp_v + tn_v) * (fn_v + tn_v))) / (counter * counter)]
  DT <- DT[, kappa := (pObs - pExp) / (1 - pExp)]
  
  best_row <- which.max(DT$kappa)
  
  if (length(best_row) > 0) {
    if (threshold) {
      return(c(DT$kappa[best_row[1]], DT$y_prob[best_row[1]]))
    } else {
      return(DT$kappa[best_row[1]])
    }
  } else {
    if (threshold) {
      return(c(-1, -1))
    } else {
      return(-1)
    }
  }
  
}