tests/testthat/test-instance_hardness.R

In themis: Extra Recipes Steps for Dealing with Unbalanced Data

test_that("errors if there isn't enough data", {
  tiny_data <- data.frame(
    x = 1:3,
    y = 1:3,
    class = factor(c("majority", "majority", "minority"))
  )

  expect_snapshot(
    error = TRUE,
    recipe(class ~ x + y, data = tiny_data) |>
      step_instance_hardness(class, neighbors = 5) |>
      prep() |>
      bake(new_data = NULL)
  )
})

test_that("basic usage", {
  rec1 <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class)

  rec1_p <- prep(rec1)

  te_xtab <- table(bake(rec1_p, new_data = circle_example)$class, useNA = "no")
  og_xtab <- table(circle_example$class, useNA = "no")

  expect_equal(sort(te_xtab), sort(og_xtab))

  expect_no_warning(prep(rec1))
})

test_that("works with a single predictor", {
  skip_if_not_installed("modeldata")

  data("hpc_data", package = "modeldata")

  expect_no_error(
    recipe(class ~ compounds, data = hpc_data) |>
      step_instance_hardness(all_outcomes()) |>
      prep() |>
      bake(NULL)
  )
})

test_that("bad data", {
  rec <- recipe(~., data = circle_example)
  # numeric check
  expect_snapshot(
    error = TRUE,
    rec |>
      step_instance_hardness(x) |>
      prep()
  )
  # Multiple variable check
  expect_snapshot(
    error = TRUE,
    rec |>
      step_instance_hardness(class, id) |>
      prep()
  )
})

test_that("errors if character are present", {
  df_char <- data.frame(
    x = factor(1:2),
    y = c("A", "A"),
    stringsAsFactors = FALSE
  )

  expect_snapshot(
    error = TRUE,
    recipe(~., data = df_char) |>
      step_instance_hardness(x) |>
      prep()
  )
})

test_that("NA in response", {
  skip_if_not_installed("modeldata")

  data("credit_data", package = "modeldata")

  expect_snapshot(
    error = TRUE,
    recipe(Job ~ Age, data = credit_data) |>
      step_instance_hardness(Job) |>
      prep()
  )
})

test_that("test tidy()", {
  rec <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class, id = "")

  rec_p <- prep(rec)

  untrained <- tibble(
    terms = "class",
    id = ""
  )

  trained <- tibble(
    terms = "class",
    id = ""
  )

  expect_equal(untrained, tidy(rec, number = 1))
  expect_equal(trained, tidy(rec_p, number = 1))
})

test_that("ratio value works when undersampling", {
  res1 <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class) |>
    prep() |>
    bake(new_data = NULL)

  res1.5 <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class, under_ratio = 1.5) |>
    prep() |>
    bake(new_data = NULL)

  expect_equal(
    as.vector(table(res1$class)),
    rep(min(table(circle_example$class)), length(table(res1$class)))
  )
  expect_equal(
    sort(as.numeric(table(res1.5$class))),
    min(table(circle_example$class)) * c(1, 1.5)
  )
})

test_that("allows multi-class", {
  skip_if_not_installed("modeldata")

  data("credit_data", package = "modeldata")
  expect_no_error(
    recipe(Home ~ Age + Income + Assets, data = credit_data) |>
      step_impute_mean(Income, Assets) |>
      step_instance_hardness(Home)
  )
})

test_that("minority classes are ignored if there is more than 1", {
  skip_if_not_installed("modeldata")

  data("penguins", package = "modeldata")
  rec1_p2 <- recipe(
    species ~ bill_length_mm + bill_depth_mm,
    data = penguins[-(1:84), ]
  ) |>
    step_impute_mean(all_predictors()) |>
    step_instance_hardness(species) |>
    prep() |>
    bake(new_data = NULL)

  expect_true(all(max(table(rec1_p2$species)) == 68))
})

test_that("factor levels are not affected by alphabet ordering or class sizes", {
  circle_example_alt_levels <- list()
  for (i in 1:4) {
    circle_example_alt_levels[[i]] <- circle_example
  }

  # Checking for forgetting levels by majority/minor switching
  for (i in c(2, 4)) {
    levels(circle_example_alt_levels[[i]]$class) <-
      rev(levels(circle_example_alt_levels[[i]]$class))
  }

  # Checking for forgetting levels by alphabetical switching
  for (i in c(3, 4)) {
    circle_example_alt_levels[[i]]$class <-
      factor(
        x = circle_example_alt_levels[[i]]$class,
        levels = rev(levels(circle_example_alt_levels[[i]]$class))
      )
  }

  for (i in 1:4) {
    rec_p <- recipe(class ~ x + y, data = circle_example_alt_levels[[i]]) |>
      step_instance_hardness(class) |>
      prep()

    expect_equal(
      levels(circle_example_alt_levels[[i]]$class), # Original levels
      rec_p$levels$class$values # New levels
    )
    expect_equal(
      levels(circle_example_alt_levels[[i]]$class), # Original levels
      levels(bake(rec_p, new_data = NULL)$class) # New levels
    )
  }
})

test_that("distance_with allows non-numeric columns to be present", {
  df_mixed <- data.frame(
    x = c(rnorm(50, 0, 1), rnorm(20, 3, 1)),
    y = c(rnorm(50, 0, 1), rnorm(20, 3, 1)),
    name = c(rep("alice", 50), rep("bob", 20)),
    class = factor(c(rep("majority", 50), rep("minority", 20)))
  )

  expect_no_error(
    recipe(class ~ ., data = df_mixed) |>
      step_instance_hardness(class, distance_with = c(x, y)) |>
      prep() |>
      bake(new_data = NULL)
  )
})

test_that("distance_with errors on non-numeric column", {
  df_mixed <- data.frame(
    x = c(1:5, 1:2),
    name = c(rep("a", 5), rep("b", 2)),
    class = factor(c(rep("majority", 5), rep("minority", 2)))
  )

  expect_snapshot(
    error = TRUE,
    recipe(class ~ ., data = df_mixed) |>
      step_instance_hardness(class, distance_with = c(x, name)) |>
      prep()
  )
})

test_that("id variables are ignored", {
  rec_id <- recipe(class ~ ., data = circle_example) |>
    update_role(id, new_role = "id") |>
    step_instance_hardness(class, under_ratio = 1) |>
    prep()

  expect_equal(ncol(bake(rec_id, new_data = NULL)), 4)
})

test_that("id variables don't turn predictors to factors", {
  rec_id <- recipe(class ~ ., data = circle_example) |>
    update_role(id, new_role = "id") |>
    step_instance_hardness(class, under_ratio = 1) |>
    prep() |>
    bake(new_data = NULL)

  expect_equal(is.double(rec_id$x), TRUE)
  expect_equal(is.double(rec_id$y), TRUE)
})

test_that("tunable", {
  rec <- recipe(~., data = mtcars) |>
    step_instance_hardness(all_predictors())
  rec_param <- tunable.step_instance_hardness(rec$steps[[1]])
  expect_equal(rec_param$name, c("under_ratio", "neighbors"))
  expect_true(all(rec_param$source == "recipe"))
  expect_true(is.list(rec_param$call_info))
  expect_equal(nrow(rec_param), 2)
  expect_equal(
    names(rec_param),
    c("name", "call_info", "source", "component", "component_id")
  )
})

test_that("distance argument accepted by step_instance_hardness()", {
  for (dist in c(
    "euclidean",
    "cosine",
    "mahalanobis",
    "manhattan",
    "chebyshev"
  )) {
    expect_no_error(
      recipe(class ~ x + y, data = circle_example) |>
        step_instance_hardness(class, distance = dist) |>
        prep() |>
        bake(new_data = NULL)
    )
  }
})

test_that("bad distance arg for step_instance_hardness()", {
  expect_snapshot(
    error = TRUE,
    bake(
      prep(step_instance_hardness(
        recipe(class ~ x + y, data = circle_example),
        class,
        distance = "L2"
      )),
      new_data = NULL
    )
  )
})

test_that("bad args", {
  expect_snapshot(
    error = TRUE,
    recipe(~., data = mtcars) |>
      step_instance_hardness(over_ratio = "yes") |>
      prep()
  )
  expect_snapshot(
    error = TRUE,
    recipe(~., data = mtcars) |>
      step_instance_hardness(neighbors = TRUE) |>
      prep()
  )
  expect_snapshot(
    error = TRUE,
    recipe(~., data = mtcars) |>
      step_instance_hardness(seed = TRUE)
  )
})

test_that("tunable is setup to works with extract_parameter_set_dials", {
  skip_if_not_installed("dials")
  rec <- recipe(~., data = mtcars) |>
    step_instance_hardness(
      all_predictors(),
      under_ratio = hardhat::tune(),
      neighbors = hardhat::tune()
    )

  params <- extract_parameter_set_dials(rec)

  expect_s3_class(params, "parameters")
  expect_identical(nrow(params), 2L)
})

test_that("unused outcome levels are skipped with a warning (#238)", {
  circle_example$class <- factor(
    circle_example$class,
    levels = c(levels(circle_example$class), "unused")
  )

  expect_snapshot(
    res <- recipe(class ~ x + y, data = circle_example) |>
      step_instance_hardness(class) |>
      prep() |>
      bake(new_data = NULL)
  )

  expect_gt(nrow(res), 0)
})

test_that("instance_hardness() works with a character `var` (#261)", {
  df <- circle_example[c("x", "y", "class")]
  df$class <- as.character(df$class)

  res <- instance_hardness(df, "class")

  expect_type(res$class, "character")
  expect_identical(sort(unique(res$class)), c("Circle", "Rest"))
  expect_identical(sum(is.na(res$class)), 0L)
})

test_that("instance_hardness() with a constant vector matches the scalar (#323)", {
  set.seed(1)
  df <- data.frame(
    x = rnorm(70),
    y = rnorm(70),
    class = factor(c(rep("a", 10), rep("b", 20), rep("c", 40)))
  )

  set.seed(2)
  res_vec <- instance_hardness(
    df,
    "class",
    under_ratio = c(a = 1.5, b = 1.5, c = 1.5)
  )
  set.seed(2)
  res_scalar <- instance_hardness(df, "class", under_ratio = 1.5)

  expect_equal(res_vec, res_scalar)
})

test_that("instance_hardness() targets a single class with a named vector (#323)", {
  set.seed(1)
  df <- data.frame(
    x = rnorm(70),
    y = rnorm(70),
    class = factor(c(rep("a", 10), rep("b", 20), rep("c", 40)))
  )

  res <- instance_hardness(df, "class", under_ratio = c(c = 2))

  expect_equal(as.numeric(table(res$class)), c(10, 20, 20))
})

test_that("step_instance_hardness() samples each class to its own target (#323)", {
  set.seed(1)
  df <- data.frame(
    x = rnorm(70),
    y = rnorm(70),
    class = factor(c(rep("a", 10), rep("b", 20), rep("c", 40)))
  )

  res <- recipe(class ~ ., data = df) |>
    step_instance_hardness(class, under_ratio = c(b = 1.5, c = 2)) |>
    prep() |>
    bake(new_data = NULL)

  expect_equal(as.numeric(table(res$class)), c(10, 15, 20))
})

test_that("step_instance_hardness() leaves a class alone when its target is on the wrong side (#323)", {
  set.seed(1)
  df <- data.frame(
    x = rnorm(70),
    y = rnorm(70),
    class = factor(c(rep("a", 10), rep("b", 20), rep("c", 40)))
  )

  res <- recipe(class ~ ., data = df) |>
    step_instance_hardness(class, under_ratio = c(a = 5)) |>
    prep() |>
    bake(new_data = NULL)

  expect_equal(as.numeric(table(res$class)), c(10, 20, 40))
})

test_that("step_instance_hardness() checks `under_ratio` names when prepped (#323)", {
  set.seed(1)
  df <- data.frame(
    x = rnorm(70),
    y = rnorm(70),
    class = factor(c(rep("a", 10), rep("b", 20), rep("c", 40)))
  )

  expect_snapshot(
    error = TRUE,
    recipe(class ~ ., data = df) |>
      step_instance_hardness(class, under_ratio = c(a = 1, potato = 1)) |>
      prep()
  )
})

# Infrastructure ---------------------------------------------------------------

test_that("bake method errors when needed non-standard role columns are missing", {
  rec <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class, skip = FALSE) |>
    add_role(class, new_role = "potato") |>
    update_role_requirements(role = "potato", bake = FALSE)

  trained <- prep(rec, training = circle_example, verbose = FALSE)

  expect_snapshot(
    error = TRUE,
    bake(trained, new_data = circle_example[, -3])
  )
})

test_that("empty printing", {
  rec <- recipe(mpg ~ ., mtcars)
  rec <- step_instance_hardness(rec)

  expect_snapshot(rec)

  rec <- prep(rec, mtcars)

  expect_snapshot(rec)
})

test_that("empty selection prep/bake is a no-op", {
  rec1 <- recipe(mpg ~ ., mtcars)
  rec2 <- step_instance_hardness(rec1)

  rec1 <- prep(rec1, mtcars)
  rec2 <- prep(rec2, mtcars)

  baked1 <- bake(rec1, mtcars)
  baked2 <- bake(rec2, mtcars)

  expect_identical(baked1, baked2)
})

test_that("empty selection tidy method works", {
  rec <- recipe(mpg ~ ., mtcars)
  rec <- step_instance_hardness(rec)

  expect <- tibble(terms = character(), id = character())

  expect_identical(tidy(rec, number = 1), expect)

  rec <- prep(rec, mtcars)

  expect_identical(tidy(rec, number = 1), expect)
})

test_that("printing", {
  rec <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class)

  expect_snapshot(print(rec))
  expect_snapshot(prep(rec))
})

test_that("0 and 1 rows data work in bake method", {
  rec <- recipe(class ~ x + y, data = circle_example) |>
    step_instance_hardness(class, skip = FALSE) |>
    prep()

  expect_identical(nrow(bake(rec, new_data = slice(circle_example, 0))), 0L)
  expect_identical(nrow(bake(rec, new_data = slice(circle_example, 1))), 1L)
})