fix(bias): use correct definition

Author: sebffischer

NEWS.md

@@ -1,4 +1,5 @@
 # mlr3measures (development version)
+* fix: Define bias measures correctly.
 
 # mlr3measures 1.0.0
 

R/regr_bias.R

@@ -2,9 +2,9 @@
 #'
 #' @details
 #' The Bias is defined as \deqn{
-#'   \frac{1}{n} \sum_{i=1}^n w_i \left( t_i - r_i \right),
+#'   \frac{1}{n} \sum_{i=1}^n w_i \left( r_i - t_i \right),
 #' }{
-#'   weighted.mean(t - r, w),
+#'   weighted.mean(r - t, w),
 #' }
 #' where \eqn{w_i} are normalized sample weights.
 #' Good predictions score close to 0.
@@ -17,7 +17,7 @@
 #' @export
 bias = function(truth, response, sample_weights = NULL, ...) {
   assert_regr(truth, response = response)
-  wmean(truth - response, sample_weights)
+  wmean(response - truth, sample_weights)
 }
 
 #' @include measures.R

R/regr_pbias.R

@@ -2,9 +2,9 @@
 #'
 #' @details
 #' The Percent Bias is defined as \deqn{
-#'   \frac{1}{n} \sum_{i=1}^n w_i \frac{\left( t_i - r_i \right)}{\left| t_i \right|},
+#'   \frac{1}{n} \sum_{i=1}^n w_i \frac{\left( r_i - t_i \right)}{\left| t_i \right|},
 #' }{
-#'   weighted.mean((t - r) / abs(t), w),
+#'   weighted.mean((r - t) / abs(t), w),
 #' }
 #' where \eqn{w_i} are normalized sample weights.
 #' Good predictions score close to 0.
@@ -20,7 +20,7 @@ pbias = function(truth, response, sample_weights = NULL, na_value = NaN, ...) {
   if (any(abs(truth) < TOL)) {
     return(na_value)
   }
-  wmean((truth - response) / abs(truth), sample_weights)
+  wmean((response - truth) / abs(truth), sample_weights)
 }
 
 #' @include measures.R

man/bias.Rd

@@ -27,9 +27,9 @@ test_that("trigger all", {
 
 test_that("tests from Metrics", {
   expect_equal(bias(1, 1), 0)
-  expect_equal(bias(c(-1, -100, 17.5), c(0, 0, 0)), mean(c(-1, -100, 17.5)))
+  expect_equal(bias(c(-1, -100, 17.5), c(0, 0, 0)), mean(c(1, 100, -17.5)))
 
-  expect_equal(pbias(c(1, 2, 3), c(1, 3, 2)), mean(c(0, -1 / 2, 1 / 3)))
+  expect_equal(pbias(c(1, 2, 3), c(1, 3, 2)), mean(c(0, 1 / 2, -1 / 3)))
   expect_equal(pbias(c(1, 2, 0), c(1, 2, 1)), NaN)
   expect_equal(pbias(0, 0), NaN)
   expect_equal(pbias(c(-1.1, 1.1), c(-1, 1)), 0)