tests/testthat/test_classif_blackboost.R

In Najah-lshanableh/R-data-mining2: Machine Learning in R

context("classif_blackboost")

test_that("classif_blackboost", {
  requirePackagesOrSkip(c("mboost", "party"), default.method = "load")

  parset.list1 = list(
    list(family = mboost::Binomial()),
    # the blackboost defaults for tree_controls needs to be passed explicitely,
    # since the defaults of party::ctree_control() differ from defaults used within blackboost
    list(family = mboost::Binomial(),
      control = mboost::boost_control(mstop = 10L),
      tree_controls = party::ctree_control(teststat = "max", testtype = "Teststatistic",
        mincriterion = 0, maxdepth = 4, savesplitstats = FALSE)),
    list(family = mboost::Binomial(link = "probit"),
      control = mboost::boost_control(mstop = 10L, nu = 0.03))
  )

  parset.list2 = list(
    list(),
    list(mstop = 10L, maxdepth = 4),
    list(Binomial.link = "probit", mstop = 10L, nu = 0.03)
  )

  old.predicts.list = list()
  old.probs.list = list()

  for (i in seq_along(parset.list1)) {
    parset = parset.list1[[i]]
    pars = list(binaryclass.formula, data = binaryclass.train)
    pars = c(pars, parset)
    set.seed(getOption("mlr.debug.seed"))
    m = do.call(mboost::blackboost, pars)
    set.seed(getOption("mlr.debug.seed"))
    old.predicts.list[[i]] = predict(m, newdata = binaryclass.test, type = "class")
    set.seed(getOption("mlr.debug.seed"))
    old.probs.list[[i]] = 1 - predict(m, newdata = binaryclass.test, type = "response")[, 1]
  }

  testSimpleParsets("classif.blackboost", binaryclass.df, binaryclass.target,
    binaryclass.train.inds, old.predicts.list, parset.list2)
  testProbParsets("classif.blackboost", binaryclass.df, binaryclass.target,
    binaryclass.train.inds, old.probs.list, parset.list2)
})


test_that("classif_blackboost probability predictions with family 'AUC' and 'AdaExp'", {
  families = list("AUC", "AdaExp")
  lapply(families, FUN = function(x){
    lrn = makeLearner("classif.blackboost", par.vals = list(family = x), predict.type = "prob")
    mod = train(lrn, binaryclass.task)
    expect_error(predict(mod, binaryclass.task), "support probabilities")
  })
})


# mlr does not support ordered factors as target yet.
# FIXME: the following two tests can be used, when they are supported
# test_that("classif_blackboost works with family PropOdds", {
#   new.binary.df = binaryclass.df
#   new.binary.df[,binaryclass.target] = as.ordered(new.binary.df[,binaryclass.target])
#   new.classif.train = new.binary.df[binaryclass.train.inds,]
#   new.classif.test = new.binary.df[binaryclass.test.inds,]
#   parset.list1 = list(family = mboost::PropOdds())
#   parset.list2 = list(family = "PropOdds")
#   old.predicts.list = list()
#   old.probs.list = list()
#   parset = parset.list1
#   pars = list(binaryclass.formula, data = new.classif.train)
#   pars = c(pars, parset)
#   set.seed(getOption("mlr.debug.seed"))
#   m = do.call(mboost::blackboost, pars)
#   set.seed(getOption("mlr.debug.seed"))
#   old.predicts.list = predict(m, newdata = new.classif.test, type = "class")
#   set.seed(getOption("mlr.debug.seed"))
#   old.probs.list = predict(m, newdata = new.classif.test, type = "response")[,1]
#
#   testSimple("classif.blackboost", new.binary.df, binaryclass.target, binaryclass.train.inds, old.predicts.list, parset.list2)
#   testProb("classif.blackboost", new.binary.df, binaryclass.target, binaryclass.train.inds, old.probs.list, parset.list2)
#
# })

# test_that("classif_blackboost with family `PropOdds` works with one observation", {
#   data = getTaskData(binaryclass.task)
#   data[,binaryclass.target] = as.ordered(data[,binaryclass.target])
#   ordered.task = makeClassifTask(data = data, target = binaryclass.target)
#   mini.data = data[1,]
#   mini.task = makeClassifTask(data = mini.data, target = binaryclass.target)
#   lrn = makeLearner("classif.blackboost", par.vals = list(family = "PropOdds"),
#     predict.type = "prob")
#   mod = train(lrn, ordered.task)
#   pred = predict(mod, mini.task)
#   orig.mod = mboost::blackboost(binaryclass.formula, data = data, family = mboost::PropOdds())
#   orig.pred = predict(orig.mod, newdata = mini.data, type = "response")
#   expect_equal(getPredictionProbabilities(pred), orig.pred[1])
# })