tests/testthat/test-num-iic.R

In yardstick: Tidy Characterizations of Model Performance

# All tests (excepted weighted ones) confirmed against the software:
# http://www.insilico.eu/coral/SOFTWARECORAL.html

test_that("Calculations are correct", {
  ex_dat <- generate_numeric_test_data()

  expect_equal(
    iic_vec(truth = ex_dat$obs, estimate = ex_dat$pred),
    0.43306222006167
  )
})

test_that("both interfaces gives the same results", {
  ex_dat <- generate_numeric_test_data()

  expect_identical(
    iic_vec(ex_dat$obs, ex_dat$pred),
    iic(ex_dat, obs, pred)[[".estimate"]],
  )
})

test_that("Calculations handles NAs", {
  ex_dat <- generate_numeric_test_data()
  na_ind <- 1:10
  ex_dat$pred[na_ind] <- NA

  expect_identical(
    iic_vec(ex_dat$obs, ex_dat$pred, na_rm = FALSE),
    NA_real_
  )

  expect_equal(
    iic_vec(truth = ex_dat$obs, estimate = ex_dat$pred),
    0.34628428506905
  )
})

test_that("Case weights calculations are correct", {
  df <- dplyr::tibble(
    truth = c(1, 2, 3, 4, 5),
    estimate = c(1, 3, 1, 3, 2),
    weight = c(1, 2, 1, 2, 0)
  )

  expect_equal(
    iic_vec(truth = df$truth, estimate = df$estimate, case_weights = df$weight),
    0.4264014327112208846415
  )
})

test_that("works with hardhat case weights", {
  solubility_test$weights <- floor(read_weights_solubility_test())
  df <- solubility_test

  imp_wgt <- hardhat::importance_weights(df$weights)
  freq_wgt <- hardhat::frequency_weights(df$weights)

  expect_no_error(
    iic_vec(df$solubility, df$prediction, case_weights = imp_wgt)
  )

  expect_no_error(
    iic_vec(df$solubility, df$prediction, case_weights = freq_wgt)
  )
})

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

test_that("iic() - result can be negative", {
  expect_equal(iic_vec(c(1, 2, 3), c(2, 1, 1)), -0.577350269189626)
})

test_that("iic() - result is NaN if truth/estimate are equivalent", {
  expect_equal(iic_vec(c(1, 2), c(1, 2)), NaN)
})

test_that("yardstick correlation warnings are thrown", {
  cnd <- rlang::catch_cnd(iic_vec(c(1, 2), c(1, 1)))
  expect_s3_class(
    cnd,
    "yardstick_warning_correlation_undefined_constant_estimate"
  )

  cnd <- rlang::catch_cnd(iic_vec(c(1, 1), c(1, 2)))
  expect_s3_class(cnd, "yardstick_warning_correlation_undefined_constant_truth")
})

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

  df <- tibble::tibble(
    truth = c(1, 2, 3, 4, 5, 6, 7),
    off = c(7, 4, 6, 2, 4, 3, 1)
  )

  # Known to produce NaN for perfect predictions
  expect_identical(
    iic_vec(df$truth, df$truth),
    NaN
  )

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