fix bug in smote when only factor features are presente

Author: berndbischl

NEWS

@@ -2,6 +2,7 @@ mlr_2.2
 - basic R learners now have new slots: name (a descriptive name of the algorithm), 
   short.name (abbreviation that can be used in plots and tables) and note 
   (notes regarding slight changes for the mlr integration of the learner and such).
+- fix a bug in "smote" when only factor features are present
 
 - new learners:
 -- classif.lqa

R/smote.R

@@ -40,6 +40,8 @@ smote = function(task, rate, nn = 5L) {
   y = data[, target]
   x = dropNamed(data, target)
   z = getMinMaxClass(y)
+  if (z$min.size < nn)
+    stopf("You cannot set nn = %i, when the minimal class has size %i!", nn, z$min.size)
   x.min = x[z$min.inds, , drop = FALSE]
   n.min = nrow(x.min)
   n.new  = round(rate * n.min) - n.min
@@ -50,14 +52,13 @@ smote = function(task, rate, nn = 5L) {
   # convert xmin to matrix, so we can handle it better in C
   # factors are integer levels
   x.min.matrix = x.min
-  if (any(is.num)) {
+  if (any(!is.num)) {
     for (i in 1:ncol(x.min.matrix)) {
       if (!is.num[i])
-        x.min.matrix[, i] = as.integer(x.min.matrix[, i])
+        x.min.matrix[, i] = as.numeric(as.integer(x.min.matrix[, i]))
     }
   }
   x.min.matrix = as.matrix(x.min.matrix)
-
   # dist matrix on smaller class, diag = 0 so we dont find x as neighnor of x
   minclass.dist = as.matrix(daisy(x.min))
   diag(minclass.dist) = NA
@@ -69,7 +70,7 @@ smote = function(task, rate, nn = 5L) {
   res = as.data.frame(res)
 
   # convert ints back to factors
-  if (any(is.num)) {
+  if (any(!is.num)) {
     for (i in 1:ncol(res)) {
       if (!is.num[i])
         res[, i] = as.factor(as.integer(res[, i]))

src/smote.c

@@ -16,10 +16,9 @@ SEXP c_smote(SEXP s_x, SEXP s_isnum, SEXP s_nn, SEXP s_res) {
     /* select a random minority obs and random neighbor */
     j_sel = runif(0, nrow_x);
     j_nn = runif(0, ncol_nn);
-	/* matrix nn contains indexes of ncol_nn nearest neighbors for each minoriy obs (= rows); 
-	// as the indexes originate from R they are one-based, so the randomly chosen index has
-    // to be subtracted by one in order to select the right row (j_nn) of zero-based matrix x
-	*/
+    /* matrix nn contains indexes of ncol_nn nearest neighbors for each minoriy obs (= rows)
+       as the indexes originate from R they are one-based, so the randomly chosen index has
+       to be subtracted by one in order to select the right row (j_nn) of zero-based matrix x */
     j_nn = nn[j_sel + j_nn * nrow_nn] - 1;
     lambda = unif_rand();
     for (R_len_t col = 0; col < ncol_x; col++) {

tests/testthat/test_imbal_smote.R

@@ -8,6 +8,14 @@ test_that("smote works",  {
   tab2 = table(df[, binaryclass.target])
   expect_equal(tab2["M"], tab1["M"])
   expect_equal(tab2["R"], tab1["R"] * 2)
+
+  # check trivial error check
+  d = data.frame(
+    x1 = rep(c("a", "b"), 3, replace = TRUE),
+    y = rep(c("a", "b"), 3, replace = TRUE)
+  )
+  task = makeClassifTask(data = d, target = "y")
+  expect_error(smote(task, rate = 2), "minimal class has size 3")
 })
 
 test_that("smote works with rate 1 (no new examples)",  {
@@ -20,6 +28,18 @@ test_that("smote works with rate 1 (no new examples)",  {
   expect_equal(tab2["R"], tab1["R"])
 })
 
+test_that("smote works with only factor fetaures",  {
+  n = 10
+  d = data.frame(
+    x1 = sample(c("a", "b"), n, replace = TRUE),
+    x2 = sample(c("a", "b"), n, replace = TRUE),
+    y = sample(c("a", "b"), n, replace = TRUE)
+  )
+  task = makeClassifTask(data = d, target = "y")
+  task2 = smote(task, rate = 1.2, nn = 2L)
+  expect_equal(task2$task.desc$size, 11)
+})
+
 test_that("smote wrapper",  {
   rdesc = makeResampleDesc("CV", iters = 2)
   lrn1 = makeLearner("classif.rpart")