tests/testthat/test-class-fall_out.R

In yardstick: Tidy Characterizations of Model Performance

test_that("Calculations are correct - two class", {
  lst <- data_altman()
  pathology <- lst$pathology

  expect_equal(
    fall_out_vec(truth = pathology$pathology, estimate = pathology$scan),
    32 / 86
  )

  expect_equal(
    fall_out_vec(truth = pathology$pathology, estimate = pathology$scan),
    1 - spec_vec(truth = pathology$pathology, estimate = pathology$scan)
  )
})

fall_out_binary <- function(data, event_level = "first") {
  1 - spec_binary(data, event_level)
}

test_that("Calculations are correct - three class", {
  multi_ex <- data_three_by_three()
  micro <- data_three_by_three_micro()

  expect_equal(
    fall_out(multi_ex, estimator = "macro")[[".estimate"]],
    macro_metric(fall_out_binary)
  )
  expect_equal(
    fall_out(multi_ex, estimator = "macro_weighted")[[".estimate"]],
    macro_weighted_metric(fall_out_binary)
  )
  expect_equal(
    fall_out(multi_ex, estimator = "micro")[[".estimate"]],
    with(micro, sum(fp) / sum(fp + tn))
  )
})

test_that("All interfaces gives the same results", {
  lst <- data_altman()
  pathology <- lst$pathology
  path_tbl <- lst$path_tbl
  path_mat <- unclass(path_tbl)

  exp <- fall_out_vec(pathology$pathology, pathology$scan)

  expect_identical(
    fall_out(path_tbl)[[".estimate"]],
    exp
  )
  expect_identical(
    fall_out(path_mat)[[".estimate"]],
    exp
  )
  expect_identical(
    fall_out(pathology, truth = pathology, estimate = scan)[[".estimate"]],
    exp
  )
})

test_that("Calculations handles NAs", {
  lst <- data_altman()
  pathology <- lst$pathology

  expect_equal(
    fall_out_vec(truth = pathology$pathology, estimate = pathology$scan_na),
    32 / 85
  )
})

test_that("Case weights calculations are correct", {
  df <- data.frame(
    truth = factor(c("x", "y", "y", "y"), levels = c("x", "y")),
    estimate = factor(c("x", "y", "y", "x"), levels = c("x", "y")),
    case_weights = c(1L, 1L, 2L, 3L)
  )

  expect_identical(
    fall_out(df, truth, estimate, case_weights = case_weights)[[".estimate"]],
    1 / 2
  )
})

test_that("work with class_pred input", {
  skip_if_not_installed("probably")

  cp_truth <- probably::as_class_pred(two_class_example$truth, which = 1)
  cp_estimate <- probably::as_class_pred(two_class_example$predicted, which = 2)

  fct_truth <- two_class_example$truth
  fct_truth[1] <- NA

  fct_estimate <- two_class_example$predicted
  fct_estimate[2] <- NA

  expect_identical(
    fall_out_vec(fct_truth, cp_estimate),
    fall_out_vec(fct_truth, fct_estimate)
  )

  expect_identical(
    fall_out_vec(fct_truth, cp_estimate, na_rm = FALSE),
    NA_real_
  )

  expect_snapshot(
    error = TRUE,
    fall_out_vec(cp_truth, cp_estimate)
  )
})

test_that("works with hardhat case weights", {
  lst <- data_altman()
  df <- lst$pathology
  imp_wgt <- hardhat::importance_weights(seq_len(nrow(df)))
  freq_wgt <- hardhat::frequency_weights(seq_len(nrow(df)))

  expect_no_error(
    fall_out_vec(df$pathology, df$scan, case_weights = imp_wgt)
  )

  expect_no_error(
    fall_out_vec(df$pathology, df$scan, case_weights = freq_wgt)
  )
})

test_that("na_rm argument check", {
  expect_snapshot(
    error = TRUE,
    fall_out_vec(1, 1, na_rm = "yes")
  )
})

test_that("`event_level = 'second'` works", {
  lst <- data_altman()
  df <- lst$pathology

  df_rev <- df
  df_rev$pathology <- stats::relevel(df_rev$pathology, "norm")
  df_rev$scan <- stats::relevel(df_rev$scan, "norm")

  expect_equal(
    fall_out_vec(df$pathology, df$scan),
    fall_out_vec(df_rev$pathology, df_rev$scan, event_level = "second")
  )
})

test_that("Binary returns `NA` with a warning when results are undefined (#98)", {
  levels <- c("a", "b")
  truth <- factor(c("a", "a"), levels = levels)
  estimate <- factor(c("a", "b"), levels = levels)

  expect_snapshot(
    out <- fall_out_vec(truth, estimate)
  )
  expect_identical(out, NA_real_)
})

test_that("Multiclass returns averaged value with warning when results undefined (#98)", {
  levels <- c("a", "b", "c", "d")

  truth <- factor(c("a", "a", "a"), levels = levels)
  estimate <- factor(c("a", "b", "b"), levels = levels)

  expect_snapshot(out <- fall_out_vec(truth, estimate))
  expect_equal(out, (0 + 2 / 3 + 0) / 3, tolerance = 0.000001)
})

test_that("`NA` is still returned if there are some undefined values but `na.rm = FALSE`", {
  levels <- c("a", "b")
  truth <- factor(c("a", "a"), levels = levels)
  estimate <- factor(c("a", NA), levels = levels)
  expect_equal(fall_out_vec(truth, estimate, na_rm = FALSE), NA_real_)
  expect_warning(fall_out_vec(truth, estimate, na_rm = FALSE), NA)
})

test_that("range values are correct", {
  direction <- metric_direction(fall_out)
  range <- metric_range(fall_out)
  perfect <- ifelse(direction == "minimize", range[1], range[2])
  worst <- ifelse(direction == "minimize", range[2], range[1])

  df <- tibble::tibble(
    truth = factor(c("A", "A", "B", "B", "B")),
    off = factor(c("B", "B", "A", "A", "A"))
  )

  expect_equal(
    fall_out_vec(df$truth, df$truth),
    perfect
  )

  if (direction == "minimize") {
    expect_gt(fall_out_vec(df$truth, df$off), perfect)
    expect_lte(fall_out_vec(df$truth, df$off), worst)
  }
  if (direction == "maximize") {
    expect_lt(fall_out_vec(df$truth, df$off), perfect)
    expect_gte(fall_out_vec(df$truth, df$off), worst)
  }
})