fix: learners can handle new factor levels (#1403)

Author: be-marc

R/assertions.R

@@ -207,7 +207,12 @@ assert_predictable = function(task, learner) {
     predict_type = fget_keys(task$col_info, i = ids, j = "type", key = "id")
     predict_levels = fget_keys(task$col_info, i = ids, j = "levels", key = "id")
 
-    ok = all(train_type == predict_type) && all(pmap_lgl(list(x = train_levels, y = predict_levels), identical))
+
+    ok = all(train_type == predict_type)
+
+    if ("new_levels" %nin% learner$properties) {
+      ok = ok && all(pmap_lgl(list(x = train_levels, y = predict_levels), identical))
+    }
 
     if (!ok) {
       stopf("Learner '%s' received task with different column info (feature type or factor level ordering) during train and predict.", learner$id)

tests/testthat/test_Learner.R

@@ -1015,3 +1015,16 @@ test_that("config error does not trigger callback", {
   l$encapsulate("evaluate", lrn("classif.featureless"))
   expect_error(l$train(tsk("iris")), regexp = "You misconfigured the learner")
 })
+
+test_that("new_levels property is working", {
+  learner = lrn("classif.featureless")
+  task = tsk("penguins")
+  learner$train(task)
+  data = task$data()
+  set(data, i = 1L, j = "island", value = "NewIsland")
+
+  expect_error(learner$predict_newdata(data), "received task with different column info")
+
+  learner$properties = c(learner$properties, "new_levels")
+  expect_prediction(learner$predict_newdata(data))
+})