tests/testthat/test-classification-label-invariance.R

test_that("classification interpolates class frequencies, not arbitrary class codes", {
  x <- rep(0:2, each = 4)
  labels <- rep(c("A", "C", "B"), each = 4)
  run <- function(levels) NNS.reg(x, factor(labels, levels = levels),
    order = "max", point.est = c(.25, .5, .75, 1.25, 1.5), plot = FALSE)
  a <- run(c("A", "B", "C")); b <- run(c("C", "A", "B"))
  expect_equal(a$class.levels[a$Point.est], b$class.levels[b$Point.est])
  expect_equal(unclass(a$Point.probabilities), b$Point.probabilities[, c("A", "B", "C")])
  expect_equal(a$Point.probabilities[2, ], c(A = .5, B = 0, C = .5))
  expect_equal(a$class.levels[a$Point.est[2]], "A")
  expect_equal(rowSums(a$Point.probabilities), rep(1, 5))
})

test_that("class cells retain counts and numeric labels keep their identity", {
  x <- factor(rep(c("left", "right"), each = 4))
  y <- c(10, 10, 10, -7, -7, -7, -7, 10)
  fit <- NNS.reg(x, y, type = "CLASS", point.est = c("left", "right"), plot = FALSE)
  expect_equal(fit$Point.est, c(10, -7))
  expect_equal(unname(unclass(fit$Point.probabilities)), rbind(c(.25, .75), c(.75, .25)))
  relabel <- ifelse(y == 10, -100, 8)
  other <- NNS.reg(x, relabel, type = "CLASS", point.est = c("left", "right"), plot = FALSE)
  expect_equal(other$Point.est, c(-100, 8))
})

test_that("stack probabilities and decoded decisions are invariant to nominal label order", {
  x <- iris[, 1:3]
  run <- function(y) NNS.stack(x, y, IVs.test = x[c(1, 52, 103), ],
    folds = 2, seed = 17, ncores = 1, status = FALSE)
  a <- run(iris$Species)
  b <- run(factor(iris$Species, levels = rev(levels(iris$Species))))
  expect_equal(a$class.levels[a$stack], b$class.levels[b$stack])
  expect_equal(unclass(a$probabilities), b$probabilities[, colnames(a$probabilities)])
  expect_equal(a$weights, b$weights)
  expect_equal(rowSums(a$probabilities), rep(1, 3))
  expect_true(all(is.finite(a$NNS.dim.red.threshold)))
  expect_true(all(is.finite(b$NNS.dim.red.threshold)))
  expect_true(a$NNS.dim.red.count %in% 1:3)
})

test_that("numeric stack outputs and custom objectives see original class values", {
  x <- iris[1:100, 1:2]
  y <- rep(c(-5, 8), each = 50)
  observed <- numeric()
  record <- function(actual) observed <<- c(observed, actual)
  loss <- as.expression(bquote({ .(record)(actual); mean(predicted == actual) }))
  fit <- NNS.stack(x, y, x[c(1, 100), ], method = 1, obj.fn = loss,
    objective = "max", folds = 2, ncores = 1, status = FALSE, seed = 9)
  expect_setequal(unique(observed), c(-5, 8))
  expect_equal(fit$stack, c(-5, 8))
})

test_that("nominal predictor level order does not alter categorical group regression", {
  y <- rep(c(2.1, 5.3, -1.8), each = 6) + rep(c(-.2, 0, .2), 6)
  labels <- rep(c("a", "b", "c"), each = 6)
  run <- function(levels) NNS.reg(factor(labels, levels = levels), y,
    point.est = c("a", "b", "c"), noise.reduction = "mean", plot = FALSE)
  a <- run(c("a", "b", "c")); b <- run(c("c", "a", "b"))
  expect_equal(a$Point.est, b$Point.est)
  expect_equal(a$Point.est, c(2.1, 5.3, -1.8))
})

test_that("both stack components canonicalize unordered predictor levels", {
  set.seed(128)
  x <- data.frame(a = rnorm(90), group = factor(rep(c("z", "a", "m"), 30)))
  reordered <- x; reordered$group <- factor(x$group, levels = c("m", "z", "a"))
  response <- x$a + rep(c(1, 2, -1), 30) + rnorm(90, sd = .1)
  run <- function(data, y) NNS.stack(data, y, IVs.test = data[1:6, ],
    method = c(1, 2), folds = 2, seed = 19, ncores = 1, status = FALSE)
  expect_equal(run(x, response), run(reordered, response))
  target <- factor(rep(c("B", "A", "C"), 30))
  expect_equal(run(x, target), run(reordered, target))
})

test_that("two-valued numeric coordinates retain distinct class cells", {
  x <- rep(c(0, 1), 20)
  y <- factor(ifelse(x == 0, "zero", "one"))
  fit <- NNS.reg(x, y, point.est = c(0, .5, 1), plot = FALSE)
  expect_equal(nrow(fit$RPM), 2L)
  expect_equal(fit$Prediction.Accuracy, 1)
  expect_equal(unname(unclass(fit$Point.probabilities)),
               rbind(c(0, 1), c(.5, .5), c(1, 0)))
  mixed <- data.frame(group = factor(x), continuous = rep(1:20, each = 2))
  cells <- NNS.reg(mixed, y, point.est = mixed[1:2, ], n.best = 1,
                   plot = FALSE, ncores = 1)
  expect_equal(cells$class.levels[cells$Point.est], c("zero", "one"))
  expect_true(all(apply(cells$RPM.probabilities, 1, max) == 1))
})

test_that("named dimension weights align in regression and nominal classification", {
  set.seed(12)
  x <- data.frame(a = rep(c(-1, 1), 20), b = rnorm(40))
  labels <- factor(ifelse(x$a > 0, "A", "B"))
  for (response in list(labels, seq_len(nrow(x)) / 10)) {
    fit <- function(w) NNS.reg(x, response, point.est = x[1:4, ],
      dim.red.method = w, plot = FALSE)
    a <- fit(c(a = 1, b = 0)); b <- fit(c(b = 0, a = 1))
    expect_equal(a$equation, b$equation)
    expect_equal(a$Point.est, b$Point.est)
    expect_equal(a$Point.probabilities, b$Point.probabilities)
    expect_error(fit(c(a = 1, unknown = 0)), "exactly match")
  }
  fit <- function(w) NNS.stack(x, labels, x[1:4, ], dim.red.method = w,
    method = c(1, 2), folds = 2, ncores = 1, seed = 1, status = FALSE)
  expect_equal(fit(c(a = 1, b = 0)), fit(c(b = 0, a = 1)))
  expect_error(fit(c(a = 1, unknown = 0)), "exactly match")
})

test_that("nominal Stack shares the factor coefficient names and fold columns", {
  x <- data.frame(group = factor(rep(c("a", "b"), 20)),
                  value = rep(1:20, each = 2))
  y <- factor(ifelse(x$group == "a", "L", "R"))
  fit <- function(w) NNS.stack(x, y, x[1:4, ], dim.red.method = w,
    method = 2, folds = 2, ncores = 1, seed = 9, status = FALSE)
  weights <- c(group__a = 1, group__b = -1, value = 0)
  a <- fit(weights); b <- fit(rev(weights))
  expect_equal(a, b)
  expect_equal(a$class.levels[a$stack], as.character(y[1:4]))
  expect_equal(a$NNS.dim.red.coefficients[[1]]$Variable,
               c(names(weights), "DENOMINATOR"))
  # A level present only in one validation row keeps its design column even
  # when absent from that fold's fitted observations.
  x$group <- as.character(x$group); x$group[1] <- "rare"
  x$group <- factor(x$group)
  rare <- NNS.stack(x, y, x[1:4, ], method = c(1, 2), folds = 2,
                    ncores = 1, seed = 9, status = FALSE)
  expect_equal(rowSums(rare$probabilities), rep(1, 4))
})

Try the NNS package in your browser

Any scripts or data that you put into this service are public.

NNS documentation built on Nov. 14, 2026, 5:09 p.m.