Merge branch 'main' into weights_reworked

.editorconfig

@@ -15,7 +15,7 @@ indent_size = 2
 indent_size = 4
 
 [*.{cpp,hpp}]
-indent_size = 4
+indent_size = 2
 
 [{NEWS.md,DESCRIPTION,LICENSE}]
 max_line_length = 80

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: mlr3proba
 Title: Probabilistic Supervised Learning for 'mlr3'
-Version: 0.6.6
+Version: 0.6.8
 Authors@R:
     c(person(given = "Raphael",
              family = "Sonabend",
@@ -64,11 +64,9 @@ Imports:
     paradox (>= 1.0.0),
     R6,
     Rcpp (>= 1.0.4),
-    survival,
-    survivalmodels (>= 0.1.12)
+    survival
 Suggests:
     bujar,
-    cubature,
     GGally,
     knitr,
     lgr,
@@ -145,6 +143,7 @@ Collate:
     'PipeOpPredRegrSurv.R'
     'PipeOpPredSurvRegr.R'
     'PipeOpProbregrCompositor.R'
+    'PipeOpResponseCompositor.R'
     'PipeOpSurvAvg.R'
     'PipeOpTaskRegrSurv.R'
     'PipeOpTaskSurvClassifDiscTime.R'
@@ -176,7 +175,6 @@ Collate:
     'histogram.R'
     'integrated_scores.R'
     'mlr3proba-package.R'
-    'partition.R'
     'pecs.R'
     'pipelines.R'
     'plot.R'

NAMESPACE

@@ -25,7 +25,6 @@ S3method(check_prediction_data,PredictionDataSurv)
 S3method(filter_prediction_data,PredictionDataSurv)
 S3method(is_missing_prediction_data,PredictionDataDens)
 S3method(is_missing_prediction_data,PredictionDataSurv)
-S3method(partition,TaskSurv)
 S3method(pecs,PredictionSurv)
 S3method(pecs,list)
 S3method(plot,LearnerSurv)
@@ -77,6 +76,7 @@ export(PipeOpPredRegrSurv)
 export(PipeOpPredSurvRegr)
 export(PipeOpPredTransformer)
 export(PipeOpProbregr)
+export(PipeOpResponseCompositor)
 export(PipeOpSurvAvg)
 export(PipeOpTaskRegrSurv)
 export(PipeOpTaskSurvClassifDiscTime)
@@ -95,7 +95,9 @@ export(as_prediction_surv)
 export(as_task_dens)
 export(as_task_surv)
 export(assert_surv)
+export(assert_surv_matrix)
 export(breslow)
+export(get_mortality)
 export(pecs)
 export(pipeline_survtoclassif_disctime)
 export(pipeline_survtoregr)

NEWS.md

@@ -1,3 +1,17 @@
+# mlr3proba 0.6.8
+
+- `Rcpp` code optimizations
+- Fixed ERV scoring to comply with `mlr3` dev version (no bugs before)
+- Skipping `survtoregr` pipelines due to bugs (to be refactored in the future)
+
+# mlr3proba 0.6.7
+
+- Deprecate `crank` to `distr` composition in `distrcompose` pipeop (only from `lp` => `distr` works now)
+- Add `get_mortality()` function (from `survivalmodels::surv_to_risk()`
+- Add Rcpp function `assert_surv_matrix()`
+- Update and simplify `crankcompose` pipeop and respective pipeline (no `response` is created anymore)
+- Add `responsecompositor` pipeline with `rmst` and `median`
+
 # mlr3proba 0.6.6
 
 - Small fixes and refactoring to the discrete-time pipeops

R/LearnerDens.R

@@ -36,14 +36,14 @@ LearnerDens = R6::R6Class("LearnerDens",
     #' @description Creates a new instance of this [R6][R6::R6Class] class.
     initialize = function(id, param_set = ps(),
       predict_types = "cdf", feature_types = character(),
-      properties = character(), data_formats = "data.table",
+      properties = character(),
       packages = character(),
       label = NA_character_,
       man = NA_character_) {
       super$initialize(
         id = id, task_type = "dens", param_set = param_set,
         predict_types = predict_types, feature_types = feature_types, properties = properties,
-        data_formats = data_formats, packages = c("mlr3proba", packages), label = label, man = man)
+        packages = c("mlr3proba", packages), label = label, man = man)
     }
   )
 )

R/MeasureSurvCindex.R

@@ -56,7 +56,7 @@
 #' library(mlr3)
 #' task = tsk("rats")
 #' learner = lrn("surv.coxph")
-#' part = partition(task) # train/test split, stratified on `status` by default
+#' part = partition(task) # train/test split
 #' learner$train(task, part$train)
 #' p = learner$predict(task, part$test)
 #'

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 problem is also 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.

R/PipeOpCrankCompositor.R

@@ -16,57 +16,36 @@
 #' ```
 #'
 #' @section Input and Output Channels:
-#' [PipeOpCrankCompositor] has one input channel named "input", which takes
-#' `NULL` during training and [PredictionSurv] during prediction.
+#' [PipeOpCrankCompositor] has one input channel named `"input"`, which takes `NULL` during training and [PredictionSurv] during prediction.
 #'
-#' [PipeOpCrankCompositor] has one output channel named "output", producing `NULL` during training
-#' and a [PredictionSurv] during prediction.
+#' [PipeOpCrankCompositor] has one output channel named `"output"`, producing `NULL` during training and a [PredictionSurv] during prediction.
 #'
-#' The output during prediction is the [PredictionSurv] from the "pred" input but with the `crank`
-#' predict type overwritten by the given estimation method.
+#' The output during prediction is the [PredictionSurv] from the input but with the `crank` predict type overwritten by the given estimation method.
 #'
 #' @section State:
 #' The `$state` is left empty (`list()`).
 #'
 #' @section Parameters:
 #' * `method` :: `character(1)` \cr
 #'    Determines what method should be used to produce a continuous ranking from the distribution.
-#'    One of `sum_haz`, `median`, `mode`, or `mean` corresponding to the
-#'    respective functions in the predicted survival distribution. Note that
-#'    for models with a proportional hazards form, the ranking implied by
-#'    `mean` and `median` will be identical (but not the value of `crank`
-#'    itself). `sum_haz` (default) uses [survivalmodels::surv_to_risk()].
-#' * `which` :: `numeric(1)`\cr
-#'    If `method = "mode"` then specifies which mode to use if multi-modal, default is the first.
-#' * `response` :: `logical(1)`\cr
-#'    If `TRUE` then the `response` predict type is estimated with the same values as `crank`.
+#'    Currently only `mort` is supported, which is the sum of the cumulative hazard, also called *expected/ensemble mortality*, see Ishwaran et al. (2008).
+#'    For more details, see [get_mortality()].
 #' * `overwrite` :: `logical(1)` \cr
-#'    If `FALSE` (default) then if the "pred" input already has a `crank`, the compositor only
-#'    composes a `response` type if `response = TRUE` and does not already exist. If `TRUE` then
-#'    both the `crank` and `response` are overwritten.
-#'
-#' @section Internals:
-#' The `median`, `mode`, or `mean` will use analytical expressions if possible but if not they are
-#' calculated using methods from [distr6]. `mean` requires \CRANpkg{cubature}.
+#'    If `FALSE` (default) and the prediction already has a `crank` prediction, then the compositor returns the input prediction unchanged.
+#'    If `TRUE`, then the `crank` will be overwritten.
 #'
 #' @seealso [pipeline_crankcompositor]
 #' @family survival compositors
 #' @examples
 #' \dontrun{
 #' if (requireNamespace("mlr3pipelines", quietly = TRUE)) {
-#'   library(mlr3)
 #'   library(mlr3pipelines)
 #'   task = tsk("rats")
 #'
-#'   learn = lrn("surv.coxph")$train(task)$predict(task)
-#'   poc = po("crankcompose", param_vals = list(method = "sum_haz"))
-#'   poc$predict(list(learn))[[1]]
-#'
-#'   if (requireNamespace("cubature", quietly = TRUE)) {
-#'     learn = lrn("surv.coxph")$train(task)$predict(task)
-#'     poc = po("crankcompose", param_vals = list(method = "sum_haz"))
-#'     poc$predict(list(learn))[[1]]
-#'   }
+#'   # change the crank prediction type of a Cox's model predictions
+#'   pred = lrn("surv.coxph")$train(task)$predict(task)
+#'   poc = po("crankcompose", param_vals = list(overwrite = TRUE))
+#'   poc$predict(list(pred))[[1L]]
 #' }
 #' }
 #' @export
@@ -77,21 +56,18 @@ PipeOpCrankCompositor = R6Class("PipeOpCrankCompositor",
     #' Creates a new instance of this [R6][R6::R6Class] class.
     initialize = function(id = "crankcompose", param_vals = list()) {
       param_set = ps(
-        method = p_fct(default = "sum_haz", levels = c("sum_haz", "mean", "median", "mode"),
-          tags = "predict"),
-        which = p_int(1L, default = 1L, tags = "predict", depends = quote(method == "mode")),
-        response = p_lgl(default = FALSE, tags = "predict"),
+        method = p_fct(default = "mort", levels = c("mort"), tags = "predict"),
         overwrite = p_lgl(default = FALSE, tags = "predict")
       )
-      param_set$set_values(method = "sum_haz", response = FALSE, overwrite = FALSE)
+      param_set$set_values(method = "mort", overwrite = FALSE)
 
       super$initialize(
         id = id,
         param_set = param_set,
         param_vals = param_vals,
         input = data.table(name = "input", train = "NULL", predict = "PredictionSurv"),
         output = data.table(name = "output", train = "NULL", predict = "PredictionSurv"),
-        packages = c("mlr3proba", "distr6")
+        packages = c("mlr3proba")
       )
     }
   ),
@@ -103,83 +79,47 @@ PipeOpCrankCompositor = R6Class("PipeOpCrankCompositor",
     },
 
     .predict = function(inputs) {
-
-      inpred = inputs[[1L]]
-
-      response = self$param_set$values$response
-      b_response = !anyMissing(inpred$response)
-      if (!length(response)) response = FALSE
-
+      pred = inputs[[1L]]
       overwrite = self$param_set$values$overwrite
-      if (!length(overwrite)) overwrite = FALSE
+      # it's impossible for a learner not to predict crank in mlr3proba,
+      # but let's check either way:
+      has_crank = !all(is.na(pred$crank))
 
-      # if crank and response already exist and not overwriting then return prediction
-      if (!overwrite && (!response || (response && b_response))) {
-        return(list(inpred))
+      if (!overwrite & has_crank) {
+        # return prediction as is
+        return(list(pred))
       } else {
-        assert("distr" %in% inpred$predict_types)
-        method = self$param_set$values$method
-        if (length(method) == 0L) method = "sum_haz"
-        if (method == "sum_haz") {
-          if (inherits(inpred$data$distr, "matrix") ||
-            !requireNamespace("survivalmodels", quietly = TRUE)) {
-            comp = survivalmodels::surv_to_risk(inpred$data$distr)
-          } else {
-            comp = as.numeric(
-              colSums(inpred$distr$cumHazard(sort(unique(inpred$truth[, 1]))))
-            )
-          }
-        } else if (method == "mean") {
-          comp = try(inpred$distr$mean(), silent = TRUE)
-          if (inherits(comp, "try-error")) {
-            requireNamespace("cubature")
-            comp = try(inpred$distr$mean(cubature = TRUE), silent = TRUE)
-          }
-          if (inherits(comp, "try-error")) {
-            comp = numeric(length(inpred$crank))
-          }
-        } else {
-          comp = switch(method,
-            median = inpred$distr$median(),
-            mode = inpred$distr$mode(self$param_set$values$which))
-        }
+        # compose crank from distr prediction
+        assert("distr" %in% pred$predict_types)
 
-        comp = as.numeric(comp)
-
-        # if crank exists and not overwriting then return predicted crank, otherwise compose
-        if (!overwrite) {
-          crank = inpred$crank
+        # get survival matrix
+        if (inherits(pred$data$distr, "array")) {
+          surv = pred$data$distr
+          if (length(dim(surv)) == 3L) {
+            # survival 3d array, extract median
+            surv = .ext_surv_mat(arr = surv, which.curve = 0.5)
+          }
         } else {
-          crank = -comp
-          # missing imputed with median
-          crank[is.na(crank)] = stats::median(crank[!is.na(crank)])
-          crank[crank == Inf] = 1e3
-          crank[crank == -Inf] = -1e3
+          stop("Distribution prediction does not have a survival matrix or array
+               in the $data$distr slot")
         }
 
-        # i) not overwriting or requesting response, and already predicted
-        if (b_response && (!overwrite || !response)) {
-          response = inpred$response
-          # ii) not requesting response and doesn't exist
-        } else if (!response) {
-          response = NULL
-          # iii) requesting response and happy to overwrite
-          # iv) requesting response and doesn't exist
-        } else {
-          response = comp
-          response[is.na(response)] = 0
-          response[response == Inf | response == -Inf] = 0
+        method = self$param_set$values$method
+        if (method == "mort") {
+          crank = get_mortality(surv)
         }
 
-        if (!anyMissing(inpred$lp)) {
-          lp = inpred$lp
-        } else {
-          lp = NULL
-        }
+        # update only `crank`
+        p = PredictionSurv$new(
+          row_ids = pred$row_ids,
+          truth = pred$truth,
+          crank = crank,
+          distr = pred$distr,
+          lp = pred$lp,
+          response = pred$response
+        )
 
-        return(list(PredictionSurv$new(
-          row_ids = inpred$row_ids, truth = inpred$truth, crank = crank,
-          distr = inpred$distr, lp = lp, response = response)))
+        return(list(p))
       }
     }
   )