Merge pull request #410 from mlr-org/rcll_note

add note for RCLL

R/MeasureSurvRCLL.R

@@ -16,6 +16,10 @@
 #' density function and \eqn{S} the survival function.
 #' RCLL is proper given that censoring and survival distribution are independent, see Rindt et al. (2022).
 #'
+#' **Note**: Even though RCLL is a proper scoring rule, the calculation of \eqn{f(t)} (which in our case is discrete, i.e. it is a *probability mass function*) for time points in the test set that don't exist in the predicted survival matrix (`distr`), results in 0 values, which are substituted by `"eps"` in our implementation, therefore skewing the result towards \eqn{-log(eps)}.
+#' This is the intractable likelihood problem discussed in Rindt et al. (2022), where the authors perform interpolation to get non-zero values for the \eqn{f(t)}.
+#' Until this is handled in `mlr3proba` some way, we advise against using this measure for model evaluation.
+#'
 #' @section Parameter details:
 #' - `na.rm` (`logical(1)`)\cr
 #' If `TRUE` (default) then removes any NAs in individual score calculations.