tests/testthat/test-ConditionalKNNSampler.R

In xplainfi: Feature Importance Methods for Global Explanations

# Tests for ConditionalKNNSampler

test_that("ConditionalKNNSampler initialization works", {
	task = tgen("friedman1")$generate(n = 100)
	sampler = ConditionalKNNSampler$new(task, k = 5L)

	expect_s3_class(sampler, "ConditionalKNNSampler")
	expect_s3_class(sampler, "ConditionalSampler")
	expect_equal(sampler$label, "k-Nearest Neighbors Conditional Sampler")
	expect_s3_class(sampler$param_set, "ParamSet")

	# Check k parameter
	expect_equal(sampler$param_set$values$k, 5L)
})

test_that("ConditionalKNNSampler works with default k", {
	task = tgen("friedman1")$generate(n = 100)
	sampler = ConditionalKNNSampler$new(task) # Default k = 5L

	expect_equal(sampler$param_set$values$k, 5L)
})

test_that("ConditionalKNNSampler sampling works", {
	task = tgen("friedman1")$generate(n = 100)
	sampler = ConditionalKNNSampler$new(task, k = 5L)

	# Marginal sampling
	expect_marginal_sampling(sampler, feature = "important1", row_ids = 1:20)

	# Conditional sampling
	expect_conditional_sampling(
		sampler,
		feature = "important2",
		conditioning_set = "important1",
		row_ids = 1:20
	)

	# Multiple features
	expect_conditional_sampling(
		sampler,
		feature = c("important2", "important3"),
		conditioning_set = "important1",
		row_ids = 1:20
	)
})

test_that("ConditionalKNNSampler sample_newdata works", {
	task = tgen("friedman1")$generate(n = 100)
	sampler = ConditionalKNNSampler$new(task, k = 5L)
	test_data = task$data(rows = 1:10)

	sampled = sampler$sample_newdata(
		feature = c("important2", "important3"),
		newdata = test_data,
		conditioning_set = "important1"
	)

	expect_sampler_output_structure(sampled, task, nrows = 10)
	expect_feature_type_consistency(sampled, task)
	expect_non_sampled_unchanged(sampled, test_data, "important1")
})

test_that("ConditionalKNNSampler handles different k values", {
	task = tgen("friedman1")$generate(n = 100)
	test_data = task$data(rows = 1:10)

	# Test with k=1 (nearest neighbor)
	sampler_k1 = ConditionalKNNSampler$new(task, k = 1L)
	expect_equal(sampler_k1$param_set$values$k, 1L)

	# Test with k=20
	sampler_k20 = ConditionalKNNSampler$new(task, k = 20L)
	expect_equal(sampler_k20$param_set$values$k, 20L)

	# Both should work
	sampled_k1 = sampler_k1$sample_newdata(
		feature = "important2",
		newdata = test_data,
		conditioning_set = "important1"
	)
	sampled_k20 = sampler_k20$sample_newdata(
		feature = "important2",
		newdata = test_data,
		conditioning_set = "important1"
	)

	expect_sampler_output_structure(sampled_k1, task, nrows = 10)
	expect_sampler_output_structure(sampled_k20, task, nrows = 10)
})

test_that("ConditionalKNNSampler handles k > n_train", {
	task = tgen("friedman1")$generate(n = 50)
	sampler = ConditionalKNNSampler$new(task, k = 100L)
	test_data = task$data(rows = 1:5)

	# Should still work (uses all available training data)
	sampled = sampler$sample_newdata(
		feature = "important2",
		newdata = test_data,
		conditioning_set = "important1"
	)

	expect_sampler_output_structure(sampled, task, nrows = 5)
})

test_that("ConditionalKNNSampler is reproducible with seed", {
	task = tgen("friedman1")$generate(n = 100)
	sampler = ConditionalKNNSampler$new(task, k = 5L)
	test_data = task$data(rows = 1:10)

	sampled1 = withr::with_seed(123, {
		sampler$sample_newdata(
			feature = "important2",
			newdata = test_data,
			conditioning_set = "important1"
		)
	})

	sampled2 = withr::with_seed(123, {
		sampler$sample_newdata(
			feature = "important2",
			newdata = test_data,
			conditioning_set = "important1"
		)
	})

	expect_identical(sampled1$important2, sampled2$important2)
})

test_that("ConditionalKNNSampler conditioning_set parameter behavior", {
	task = tgen("friedman1")$generate(n = 100)
	test_conditioning_set_behavior(ConditionalKNNSampler, task, k = 5L)
})

test_that("ConditionalKNNSampler preserves feature types", {
	skip_if_not_installed("gower")
	test_sampler_feature_types(ConditionalKNNSampler, k = 5L)
})

test_that("ConditionalKNNSampler works with categorical features using Gower distance", {
	skip_if_not_installed("gower")

	task = tsk("penguins")
	data = task$data()
	sampler = ConditionalKNNSampler$new(task, k = 5L)
	test_data = task$data(rows = 1:10)

	# Sample conditioning on categorical feature
	sampled = sampler$sample_newdata(
		feature = "bill_length",
		newdata = test_data,
		conditioning_set = "island"
	)

	expect_sampler_output_structure(sampled, task, nrows = 10)
	expect_non_sampled_unchanged(sampled, test_data, "island")

	# Sampled values should come from training data
	expect_true(all(sampled$bill_length %in% data$bill_length))
})

test_that("ConditionalKNNSampler works with mixed numeric and categorical conditioning", {
	skip_if_not_installed("gower")

	task = tsk("penguins")
	sampler = ConditionalKNNSampler$new(task, k = 5L)
	test_data = task$data(rows = 1:10)

	# Sample conditioning on both numeric and factor features
	sampled = sampler$sample_newdata(
		feature = "bill_length",
		newdata = test_data,
		conditioning_set = c("island", "body_mass")
	)

	expect_sampler_output_structure(sampled, task, nrows = 10)
	expect_non_sampled_unchanged(sampled, test_data, c("island", "body_mass"))
})