Merge pull request #160 from bblodfon/uni_cox

Add Univariate cox filter

Author: sebffischer

DESCRIPTION

@@ -8,7 +8,9 @@ Authors@R: c(
            comment = c(ORCID = "0000-0001-9754-0393")),
     person("Bernd", "Bischl", , "bernd_bischl@gmx.net", role = "aut",
            comment = c(ORCID = "0000-0001-6002-6980")),
-    person("Martin", "Binder", , "mlr.developer@mb706.com", role = "aut")
+    person("Martin", "Binder", , "mlr.developer@mb706.com", role = "aut"),
+    person("John", "Zobolas", , "bblodfon@gmail.com", role = "aut",
+           comment = c(ORCID = "0000-0002-3609-8674"))
   )
 Description: Extends 'mlr3' with filter methods for feature selection.
     Besides standalone filter methods built-in methods of any
@@ -40,6 +42,7 @@ Suggests:
     mlr3pipelines,
     praznik,
     rpart,
+    survival,
     testthat (>= 3.0.0),
     withr
 Config/testthat/edition: 3
@@ -71,6 +74,7 @@ Collate:
     'FilterPermutation.R'
     'FilterRelief.R'
     'FilterSelectedFeatures.R'
+    'FilterUnivariateCox.R'
     'FilterVariance.R'
     'bibentries.R'
     'flt.R'

NAMESPACE

@@ -23,6 +23,7 @@ export(FilterPerformance)
 export(FilterPermutation)
 export(FilterRelief)
 export(FilterSelectedFeatures)
+export(FilterUnivariateCox)
 export(FilterVariance)
 export(as.data.table)
 export(flt)

NEWS.md

@@ -1,7 +1,9 @@
 # mlr3filters 0.7.1-9000
 
-* Parameter value `na.rm` is properly initialized to `TRUE` (Thanks to @bblodfon)
-* Bugfix: property `missings` is now set correctly for `FilterFindCorrelation`
+- Fixed issue with `FilterPerformance` where the arg `measure` wasn't passed on
+- Added `FilterUnivariateCox` (thanks to @bblodfon)
+- Parameter value `na.rm` is properly initialized to `TRUE` (thanks to @bblodfon)
+- Bugfix: property `missings` is now set correctly for `FilterFindCorrelation`
 
 # mlr3filters 0.7.1
 

R/FilterPerformance.R

@@ -80,7 +80,8 @@ FilterPerformance = R6Class("FilterPerformance",
 
       perf = map_dbl(fn, function(x) {
         task$col_roles$feature = x
-        resample(task, self$learner, self$resampling, clone = character())$aggregate()
+        resample(task, self$learner, self$resampling, clone = character())$
+          aggregate(measures = self$measure)
       })
 
       if (self$measure$minimize) {

R/FilterUnivariateCox.R

@@ -0,0 +1,68 @@
+#' @title Univariate Cox Survival Filter
+#'
+#' @name mlr_filters_univariate_cox
+#'
+#' @description Calculates scores for assessing the relationship between
+#' individual features and the time-to-event outcome (right-censored survival
+#' data) using a univariate Cox proportional hazards model.
+#' The goal is to determine which features have a statistically significant
+#' association with the event of interest, typically in the context of clinical
+#' or biomedical research.
+#'
+#' This filter fits a [Cox Proportional Hazards][survival::coxph()] model using
+#' each feature independently and extracts the \eqn{p}-value that quantifies the
+#' significance of the feature's impact on survival. The filter value is
+#' `-log10(p)` where `p` is the \eqn{p}-value. This transformation is necessary
+#' to ensure numerical stability for very small \eqn{p}-values. Also higher
+#' values denote more important features. The filter works only for numeric
+#' features so please ensure that factor variables are properly encoded, e.g.
+#' using [PipeOpEncode][mlr3pipelines::PipeOpEncode].
+#'
+#' @family Filter
+#' @include Filter.R
+#' @template seealso_filter
+#' @export
+#' @examples
+#'
+#' filter = flt("univariate_cox")
+#' filter
+#'
+FilterUnivariateCox = R6Class("FilterUnivariateCox",
+  inherit = Filter,
+  public = list(
+    #' @description Create a FilterUnivariateCox object.
+    initialize = function() {
+      super$initialize(
+        id = "surv.univariate_cox",
+        packages = "survival",
+        param_set = ps(),
+        feature_types = c("integer", "numeric", "logical"),
+        task_types = "surv",
+        label = "Univariate Cox Survival Score",
+        man = "mlr3filters::mlr_filters_univariate_cox"
+      )
+    }
+  ),
+
+  private = list(
+    .calculate = function(task, nfeat) {
+      features = task$feature_names
+      targets  = task$data(cols = task$target_names)
+
+      scores = map_dbl(features, function(feature) {
+        model = invoke(
+          survival::coxph,
+          formula = task$formula(rhs = feature),
+          data = cbind(task$data(cols = feature), targets)
+        )
+        pval = summary(model)$coefficients[, "Pr(>|z|)"]
+        -log10(pval) # smaller p-values => larger scores
+      })
+
+      set_names(scores, features)
+    }
+  )
+)
+
+#' @include mlr_filters.R
+mlr_filters$add("univariate_cox", FilterUnivariateCox)

README.Rmd

@@ -102,9 +102,9 @@ Please open an issue so we can add your package.
 Some learners need to have their variable importance measure "activated" during learner creation.
 For example, to use the "impurity" measure of Random Forest via the {ranger} package:
 
-```{r }
+```{r}
 task = tsk("iris")
-lrn = lrn("classif.ranger")
+lrn = lrn("classif.ranger", seed = 42)
 lrn$param_set$values = list(importance = "impurity")
 
 filter = flt("importance", learner = lrn)

README.md

@@ -84,6 +84,7 @@ head(as.data.table(filter$calculate(task)))
 | permutation       | Permutation Score                                        | Universal      | Logical, Integer, Numeric, Character, Factor, Ordered, POSIXct |                                                                                                                  |
 | relief            | RELIEF                                                   | Classif & Regr | Integer, Numeric, Factor, Ordered                              | [FSelectorRcpp](https://cran.r-project.org/package=FSelectorRcpp)                                                |
 | selected_features | Embedded Feature Selection                               | Universal      | Logical, Integer, Numeric, Character, Factor, Ordered, POSIXct |                                                                                                                  |
+| univariate_cox    | Univariate Cox Survival Score                            | Surv           | Integer, Numeric, Logical                                      | [survival](https://cran.r-project.org/package=survival)                                                          |
 | variance          | Variance                                                 | Universal      | Integer, Numeric                                               | stats                                                                                                            |
 
 ### Variable Importance Filters
@@ -108,7 +109,7 @@ Random Forest via the {ranger} package:
 
 ``` r
 task = tsk("iris")
-lrn = lrn("classif.ranger")
+lrn = lrn("classif.ranger", seed = 42)
 lrn$param_set$values = list(importance = "impurity")
 
 filter = flt("importance", learner = lrn)