predict ordinal factors from ordinal regression models (#1217)

Author: topepo

NEWS.md

@@ -25,6 +25,8 @@
 
 ## Bug Fixes
 
+* Make sure that parsnip does not convert ordered factor predictions to be unordered.
+
 * Ensure that `knit_engine_docs()` has the required packages installed (#1156).
 
 * Fixed bug where some models fit using `fit_xy()` couldn't predict (#1166).

R/fit.R

@@ -87,6 +87,7 @@
 #' \itemize{
 #'   \item \code{lvl}: If the outcome is a factor, this contains
 #'    the factor levels at the time of model fitting.
+#'   \item \code{ordered}: If the outcome is a factor, was it an ordered factor?
 #'   \item \code{spec}: The model specification object
 #'    (\code{object} in the call to \code{fit})
 #'   \item \code{fit}: when the model is executed without error,

R/fit_helpers.R

@@ -40,7 +40,8 @@ form_form <-
     fit_call <- make_form_call(object, env = env)
 
     res <- list(
-      lvl = y_levels,
+      lvl = y_levels$lvl,
+      ordered = y_levels$ordered,
       spec = object
     )
 
@@ -98,7 +99,7 @@ xy_xy <- function(object,
 
   fit_call <- make_xy_call(object, target, env, call)
 
-  res <- list(lvl = levels(env$y), spec = object)
+  res <- list(lvl = levels(env$y), ordered = is.ordered(env$y), spec = object)
 
   time <- proc.time()
   res$fit <- eval_mod(

R/misc.R

@@ -260,9 +260,12 @@ convert_arg <- function(x) {
 
 levels_from_formula <- function(f, dat) {
   if (inherits(dat, "tbl_spark")) {
-    res <- NULL
+    res <- list(lvls = NULL, ordered = FALSE)
   } else {
-    res <- levels(eval_tidy(rlang::f_lhs(f), dat))
+    res <- list()
+    y_data <- eval_tidy(rlang::f_lhs(f), dat)
+    res$lvls <- levels(y_data)
+    res$ordered <- is.ordered(y_data)
   }
   res
 }

R/predict_class.R

@@ -41,14 +41,16 @@ predict_class.model_fit <- function(object, new_data, ...) {
 
   # coerce levels to those in `object`
   if (is.vector(res) || is.factor(res)) {
-    res <- factor(as.character(res), levels = object$lvl)
+    res <- factor(as.character(res), levels = object$lvl, ordered = object$ordered)
   } else {
     if (!inherits(res, "tbl_spark")) {
       # Now case where a parsnip model generated `res`
       if (is.data.frame(res) && ncol(res) == 1 && is.factor(res[[1]])) {
         res <- res[[1]]
       } else {
-        res$values <- factor(as.character(res$values), levels = object$lvl)
+        res$values <- factor(as.character(res$values),
+                             levels = object$lvl,
+                             ordered = object$ordered)
       }
     }
   }

man/fit.Rd

@@ -43,6 +43,50 @@ test_that('classification predictions', {
                c(".pred_high", ".pred_low"))
 })
 
+
+test_that('ordinal classification predictions', {
+  skip_if_not_installed("modeldata")
+  skip_if_not_installed("rpart")
+
+  set.seed(382)
+  dat_tr <-
+    modeldata::sim_multinomial(
+      200,
+      ~  -0.5    +  0.6 * abs(A),
+      ~ ifelse(A > 0 & B > 0, 1.0 + 0.2 * A / B, - 2),
+      ~ -0.6 * A + 0.50 * B -  A * B) %>%
+    dplyr::mutate(class = as.ordered(class))
+  dat_te <-
+    modeldata::sim_multinomial(
+      5,
+      ~  -0.5    +  0.6 * abs(A),
+      ~ ifelse(A > 0 & B > 0, 1.0 + 0.2 * A / B, - 2),
+      ~ -0.6 * A + 0.50 * B -  A * B) %>%
+    dplyr::mutate(class = as.ordered(class))
+
+  ###
+
+  mod_f_fit <-
+    decision_tree() %>%
+    set_mode("classification") %>%
+    fit(class ~ ., data = dat_tr)
+  expect_true("ordered" %in% names(mod_f_fit))
+  mod_f_pred <- predict(mod_f_fit, dat_te)
+  expect_true(is.ordered(mod_f_pred$.pred_class))
+
+  ###
+
+  mod_xy_fit <-
+    decision_tree() %>%
+    set_mode("classification") %>%
+    fit_xy(x = dat_tr %>% dplyr::select(-class), dat_tr$class)
+
+  expect_true("ordered" %in% names(mod_xy_fit))
+  mod_xy_pred <- predict(mod_xy_fit, dat_te)
+  expect_true(is.ordered(mod_f_pred$.pred_class))
+})
+
+
 test_that('non-standard levels', {
   expect_true(is_tibble(predict(lr_fit, new_data = class_dat[1:5,-1])))
   expect_true(is.factor(parsnip:::predict_class.model_fit(lr_fit, new_data = class_dat[1:5,-1])))