From dcda3db48da3a56c3a898d1476ad7c44984b8d2d Mon Sep 17 00:00:00 2001
From: taekop <taekop@naver.com>
Date: Tue, 29 Sep 2026 11:32:26 +0900
Subject: [PATCH] Fix brier_class() case weight normalization

---
 NEWS.md                                             |  4 ++++
 R/prob-brier_class.R                                |  8 --------
 tests/testthat/test-orderedprob-ranked_prob_score.R |  2 +-
 tests/testthat/test-prob-brier_class.R              | 12 +++++++++++-
 4 files changed, 16 insertions(+), 10 deletions(-)

diff --git a/NEWS.md b/NEWS.md
index 5fe2d362..faaa98ec 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,5 +1,9 @@
 # yardstick (development version)
 
+## Bug fixes
+
+* `brier_class()` now computes the correct result with `case_weights` when the weights are on very different scales, rather than silently renormalizing them with a softmax (#633).
+
 # yardstick 1.4.0
 
 ## Breaking Changes
diff --git a/R/prob-brier_class.R b/R/prob-brier_class.R
index b2a51dd8..f11b3b8e 100644
--- a/R/prob-brier_class.R
+++ b/R/prob-brier_class.R
@@ -182,14 +182,6 @@ brier_ind <- function(truth, estimate, case_weights = NULL) {
   resids <- resids[not_missing, , drop = FALSE]
   case_weights <- case_weights[not_missing]
 
-  # Normalize weights (in case negative weights)
-  # subtracting max to avoid Inf in calculations
-  #  exp(x - max(x)) / sum(exp(x - max(x)))
-  #  = (exp(x) / exp(max(x))) / (sum(exp(x)) / exp(max(x)))
-  #  = exp(x) / sum(exp(x))
-  case_weights <- case_weights - max(case_weights)
-  case_weights <- exp(case_weights) / sum(exp(case_weights))
-
   res <- sum(resids * case_weights) / (2 * sum(case_weights))
   res
 }
diff --git a/tests/testthat/test-orderedprob-ranked_prob_score.R b/tests/testthat/test-orderedprob-ranked_prob_score.R
index 0452d26f..3537eaff 100644
--- a/tests/testthat/test-orderedprob-ranked_prob_score.R
+++ b/tests/testthat/test-orderedprob-ranked_prob_score.R
@@ -52,7 +52,7 @@ test_that("Case weights calculations are correct", {
   hpc_cv$obs <- as.ordered(hpc_cv$obs)
   hpc_cv$weights <- rep(c(0, 1), c(100, nrow(hpc_cv) - 100))
   # With orf:::rps(as.matrix(hpc_cv[-(1:100), 3:6]), hpc_cv$obs[-(1:100)])
-  hpc_exp <- 0.0868733
+  hpc_exp <- 0.0875957
 
   expect_equal(
     yardstick:::ranked_prob_score_vec(
diff --git a/tests/testthat/test-prob-brier_class.R b/tests/testthat/test-prob-brier_class.R
index 784e60bb..767571c3 100644
--- a/tests/testthat/test-prob-brier_class.R
+++ b/tests/testthat/test-prob-brier_class.R
@@ -114,12 +114,22 @@ test_that("doesn't produce NaN with high valued case weights (#614)", {
   wts <- rep(1, (nrow(df)))
   wts_high <- wts * 5000
 
-  expect_identical(
+  expect_equal(
     brier_class_vec(df$truth, df$Class1, case_weights = wts),
     brier_class_vec(df$truth, df$Class1, case_weights = wts_high)
   )
 })
 
+test_that("case weights match uncounted data with widely spread weights (#633)", {
+  hpc_wts <- hpc_cv
+  hpc_wts$wts <- rep(1:10, length.out = nrow(hpc_wts))
+
+  expect_equal(
+    brier_class(hpc_wts, obs, VF:L, case_weights = wts)[[".estimate"]],
+    brier_class(tidyr::uncount(hpc_wts, wts), obs, VF:L)[[".estimate"]]
+  )
+})
+
 test_that("`event_level = 'second'` works", {
   df <- two_class_example