BOT: Fix #763: Warn when binary factor levels are in counterintuitive order

R/metrics-binary.R

@@ -19,6 +19,32 @@ assert_input_binary <- function(observed, predicted) {
   assert_factor(observed, n.levels = 2, min.len = 1)
   assert_numeric(predicted, lower = 0, upper = 1)
   assert_dims_ok_point(observed, predicted)
+
+  # Warn if factor levels appear to be in counterintuitive order.
+  # Predictions represent P(outcome = highest factor level). If the levels
+  # are e.g. c("1", "0"), the highest level is "0", meaning predictions are
+  # interpreted as P(outcome = "0"), which is almost certainly unintended.
+  lvls <- levels(observed)
+  counterintuitive <- FALSE
+  if (setequal(lvls, c("0", "1")) && lvls[1] == "1") {
+    counterintuitive <- TRUE
+  } else if (setequal(lvls, c("TRUE", "FALSE")) && lvls[1] == "TRUE") {
+    counterintuitive <- TRUE
+  }
+  if (counterintuitive) {
+    #nolint start: keyword_quote_linter
+    cli_warn(c(
+      "!" = "Factor levels of {.var observed} appear to be in
+             counterintuitive order: {.val {lvls}}.",
+      "i" = "Predictions will be interpreted as the probability of
+             observing {.val {lvls[2]}} (the highest factor level).",
+      "i" = "If this is not intended, consider reordering the factor levels,
+             e.g. {.code factor(observed, levels = c({.val {lvls[2]}},
+             {.val {lvls[1]}}))}"
+    ))
+    #nolint end
+  }
+
   return(invisible(NULL))
 }
 

tests/testthat/test-class-forecast-binary.R

@@ -47,6 +47,27 @@
 
 
 
+test_that("as_forecast_binary() warns when data has reversed 0/1 factor levels", {
+  dt <- data.table(
+    model = "m1",
+    id = 1:4,
+    observed = factor(c(0, 1, 1, 0), levels = c("1", "0")),
+    predicted = c(0.1, 0.9, 0.8, 0.2)
+  )
+  expect_warning(
+    as_forecast_binary(dt),
+    "counterintuitive"
+  )
+})
+
+test_that("score() produces correct results with standard 0/1 factor levels", {
+  # example_binary has standard levels c("0", "1"), should not warn about levels
+  expect_no_warning(
+    suppressMessages(score(as_forecast_binary(example_binary)))
+  )
+})
+
+
 # ==============================================================================
 # score.forecast_binary() # nolint: commented_code_linter
 # ==============================================================================

tests/testthat/test-metrics-binary.R

@@ -106,6 +106,64 @@
 })
 
 
+# ==============================================================================
+# Test factor level order warning
+# ==============================================================================
+test_that("assert_input_binary() warns when 0/1 factor levels are in counterintuitive order", {
+  observed_rev <- factor(c(0, 1, 1, 0, 1), levels = c("1", "0"))
+  predicted_rev <- c(0.1, 0.9, 0.8, 0.2, 0.7)
+  expect_warning(
+    assert_input_binary(observed_rev, predicted_rev),
+    "counterintuitive"
+  )
+})
+
+test_that("assert_input_binary() does not warn for standard 0/1 level order", {
+  observed_std <- factor(c(0, 1, 1, 0, 1), levels = c("0", "1"))
+  predicted_std <- c(0.1, 0.9, 0.8, 0.2, 0.7)
+  expect_no_warning(assert_input_binary(observed_std, predicted_std))
+})
+
+test_that("assert_input_binary() does not warn for non-numeric factor levels", {
+  observed_ab <- factor(c("a", "b", "b", "a"), levels = c("a", "b"))
+  predicted_ab <- c(0.3, 0.7, 0.6, 0.4)
+  expect_no_warning(assert_input_binary(observed_ab, predicted_ab))
+})
+
+test_that("brier_score() produces different results with reversed factor levels", {
+  observed_correct <- factor(c(0, 1, 1, 0), levels = c("0", "1"))
+  observed_reversed <- factor(c(0, 1, 1, 0), levels = c("1", "0"))
+  predicted_bs <- c(0.1, 0.9, 0.8, 0.2)
+
+  scores_correct <- brier_score(observed_correct, predicted_bs)
+  expect_equal(scores_correct, c(0.01, 0.01, 0.04, 0.04)) # nolint: expect_identical_linter
+
+  expect_warning(
+    scores_reversed <- brier_score(observed_reversed, predicted_bs),
+    "counterintuitive"
+  )
+  expect_false(all(scores_correct == scores_reversed))
+})
+
+test_that("logs_binary() warns with reversed 0/1 factor levels", {
+  observed_reversed <- factor(c(0, 1, 1, 0), levels = c("1", "0"))
+  predicted_lb <- c(0.1, 0.9, 0.8, 0.2)
+  expect_warning(
+    logs_binary(observed_reversed, predicted_lb),
+    "counterintuitive"
+  )
+})
+
+test_that("assert_input_binary() warns for TRUE/FALSE levels in counterintuitive order", {
+  observed_tf <- factor(c(TRUE, FALSE, TRUE), levels = c("TRUE", "FALSE"))
+  predicted_tf <- c(0.8, 0.2, 0.9)
+  expect_warning(
+    assert_input_binary(observed_tf, predicted_tf),
+    "counterintuitive"
+  )
+})
+
+
 # ==============================================================================
 # Test Binary Metrics
 # ==============================================================================