fix another bug in roc/prc vis (fixes #79)

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: mlr3viz
 Title: Visualizations for 'mlr3'
-Version: 0.5.4
+Version: 0.5.5
 Authors@R:
     c(person(given = "Michel",
              family = "Lang",

NEWS.md

@@ -1,7 +1,10 @@
-# mlr3viz 0.5.3.9000
+# mlr3viz 0.5.5
 
-- Fixed a bug for ROC- and Precision-recall-curves (#72, #75).
+- Fixed another bug for ROC- and Precision-recall-curves (#79).
+
+# mlr3viz 0.5.4
 
+- Fixed a bug for ROC- and Precision-recall-curves (#72, #75).
 
 # mlr3viz 0.5.3
 

R/as_precrec.R

@@ -27,7 +27,7 @@ roc_data = function(prediction) {
   }
 
   data.table(
-    scores = prediction$prob[, 1L],
+    scores = prediction$prob[, 1L, drop = TRUE],
     labels = prediction$truth
   )
 }
@@ -42,7 +42,7 @@ as_precrec.PredictionClassif = function(object) { # nolint
     scores = data$scores,
     labels = data$labels,
     dsids = 1L,
-    posclass = levels(data$labels)[1L]
+    posclass = levels(object$truth)[1L]
   )
 }
 
@@ -53,10 +53,12 @@ as_precrec.ResampleResult = function(object) { # nolint
   require_namespaces("precrec")
   predictions = object$predictions()
   data = transpose_list(map(predictions, roc_data))
+
   precrec::mmdata(
-    scores = data$scores, labels = data$labels,
+    scores = data$scores,
+    labels = data$labels,
     dsids = seq_along(predictions),
-    posclass = levels(data$labels)[1L]
+    posclass = object$task$positive
   )
 }
 
@@ -82,10 +84,10 @@ as_precrec.BenchmarkResult = function(object) { # nolint
   lrns = unique(scores$learner_id)
   iters = unique(scores$iteration)
   precrec::mmdata(
-    data$scores,
-    data$labels,
+    scores = data$scores,
+    labels = data$labels,
     dsids = iters,
     modnames = lrns,
-    posclass = levels(data$labels)[1L]
+    posclass = object$tasks$task[[1L]]$positive
   )
 }

tests/testthat/test_BenchmarkResult.R

@@ -39,4 +39,8 @@ test_that("holdout roc plot (#54)", {
   bmr = benchmark(design)
   p = autoplot(bmr, type = "roc")
   expect_true(is.ggplot(p))
+
+  # roc is not inverted?
+  tab = as.data.table(precrec::auc(precrec::evalmod(as_precrec(bmr))))
+  expect_number(tab[modnames == "classif.rpart" & curvetypes == "ROC", aucs], lower = 0.5)
 })

tests/testthat/test_PredictionClassif.R

@@ -1,8 +1,8 @@
-test_that("autoplot.PredictionClassif", {
-  task = mlr3::tsk("sonar")
-  learner = mlr3::lrn("classif.rpart", predict_type = "prob")$train(task)
-  prediction = learner$predict(task)
+task = mlr3::tsk("sonar")
+learner = mlr3::lrn("classif.rpart", predict_type = "prob")$train(task)
+prediction = learner$predict(task)
 
+test_that("autoplot.PredictionClassif", {
   p = autoplot(prediction, type = "stacked")
   expect_true(is.ggplot(p))
 
@@ -15,3 +15,9 @@ test_that("autoplot.PredictionClassif", {
   p = autoplot(prediction, type = "threshold")
   expect_true(is.ggplot(p))
 })
+
+test_that("roc is not inverted", {
+  skip_if_not_installed("precrec")
+  tab = as.data.table(precrec::auc(precrec::evalmod(as_precrec(prediction))))
+  expect_numeric(tab[curvetypes == "ROC", aucs], len = 1L, lower = 0.5)
+})

tests/testthat/test_ResampleResult.R

@@ -72,3 +72,11 @@ test_that("autoplot ResampleResult type=prediction", {
     regexp = "Plot learner prediction only works with one or two features for
 regression!")
 })
+
+
+test_that("roc is not inverted", {
+  autoplot(rr, type = "roc")
+  skip_if_not_installed("precrec")
+  tab = as.data.table(precrec::auc(precrec::evalmod(as_precrec(rr))))
+  expect_number(mean(tab[curvetypes == "ROC", aucs]), lower = 0.5)
+})