add borda score

R/EnsembleFSResult.R

@@ -126,18 +126,26 @@ EnsembleFSResult = R6Class("EnsembleFSResult",
     #' The weights used are equal to the performance scores of each voter/model (or the inverse scores if the measure is minimized).
     #' The un-weighted methods use same weights for all voters (equal to 1).
     #'
+    #' Note that some methods output a feature ranking instead of a score per feature.
+    #' Therefore we also calculate **Borda's score**:
+    #' \eqn{s_{borda} = (p-i)/(p-1)}, where \eqn{p} is the total number of features, and \eqn{i} is the feature ranking.
+    #' So the best feature gets a borda score of \eqn{1} and the worst-ranked feature a borda score of \eqn{0}.
+    #' This score is method-agnostic, i.e. it can be used to compare the feature rankings across different methods.
+    #'
     #' The following methods are currently supported:
     #'
     #' - `"av"|"av_weighted"` (approval voting) selects the candidates that have the highest approval score, i.e. the features that appear the most often.
     #' This is the default feature ranking method.
-    #' - `"sav"|"sav_weighted"` (satisfaction approval voting) selects the candidates that have a higher satisfaction score,in proportion to the size of the voters approval sets.
+    #' - `"sav"|"sav_weighted"` (satisfaction approval voting) selects the candidates that have a higher satisfaction score, in proportion to the size of the voters approval sets.
     #' Voters who approve more candidates contribute a lesser score to the individual approved candidates.
     #'
     #' @param method (`character(1)`)\cr
     #' The method to calculate the feature ranking.
     #'
     #' @return A [data.table::data.table] listing all the features, ordered by decreasing scores (depends on the `"method"`).
     #' An extra column `"norm_score"` is produced for methods for which the original scores (i.e. approval counts in the case of approval voting) can be normalized and interpreted as **selection probabilities**, see Meinshausen et al. (2010).
+    #' The `"borda_score"` column is always included to incorporate feature ranking methods that don't output per-feature scores but only rankings.
+    #'
     feature_ranking = function(method = "av") {
       assert_choice(method, choices = c("av", "av_weighted", "sav", "sav_weighted"))
 

R/voting_methods.R

@@ -1,9 +1,10 @@
 # some collection of voting methods for feature ranking
 
-# Parameters:
 # @param voters list of feature vectors (best features, each a subset of "candidates")
 # @param candidates vector with ALL features
 # @param weights vector of weights, 1-1 correspondence with voters
+# @return data.table with 4 columns: features (mandatory), score, norm_score, borda_score (mandatory). Always features are ordered with decreasing `score` (or descreasing
+# `borda_score` if a method returns only a ranking).
 
 approval_voting = function(voters, candidates, weights) {
   # faster R version in case of equal weights
@@ -31,10 +32,11 @@ approval_voting = function(voters, candidates, weights) {
     setorderv(res, cols = "score", order = -1)
   }
 
-  res
+  res[, borda_score := (nrow(res) - .I) / (nrow(res) - 1)]
 }
 
 satisfaction_approval_voting = function(voters, candidates, weights) {
   res = as.data.table(SAV_rcpp(voters, candidates, weights))
   setorderv(res, cols = "score", order = -1)
+  res[, borda_score := (nrow(res) - .I) / (nrow(res) - 1)]
 }