parttree (version 0.1.1)

parttree: Convert a decision tree into a data frame of partition coordinates

Description

Extracts the terminal leaf nodes of a decision tree that contains no more that two numeric predictor variables. These leaf nodes are then converted into a data frame, where each row represents a partition (or leaf or terminal node) that can easily be plotted in 2-D coordinate space.

Usage

parttree(tree, keep_as_dt = FALSE, flip = FALSE)

Value

A data frame comprising seven columns: the leaf node, its path, a set of rectangle limits (i.e., xmin, xmax, ymin, ymax), and a final column corresponding to the predicted value for that leaf.

Arguments

tree

An rpart.object or alike. This includes compatible classes from the mlr3 and tidymodels frontends, or the constparty class inheriting from party.

keep_as_dt

Logical. The function relies on data.table for internal data manipulation. But it will coerce the final return object into a regular data frame (default behavior) unless the user specifies TRUE.

flip

Logical. Should we flip the "x" and "y" variables in the return data frame? The default behaviour is for the first split variable in the tree to take the "y" slot, and any second split variable to take the "x" slot. Setting to TRUE switches these around.

Note: This argument is primarily useful when it passed via geom_parttree to ensure correct axes orientation as part of a ggplot2 visualization (see geom_parttree Examples). We do not expect users to call parttree(..., flip = TRUE) directly. Similarly, to switch axes orientation for the native (base graphics) plot.parttree method, we recommend calling plot(..., flip = TRUE) rather than flipping the underlying parttree object.

See Also

plot.parttree, geom_parttree, rpart, ctree partykit::ctree.

Examples

library("parttree")
data.table::setDTthreads(2)
#
## rpart trees

library("rpart")
rp = rpart(Kyphosis ~ Start + Age, data = kyphosis)

# A parttree object is just a data frame with additional attributes
(rp_pt = parttree(rp))
attr(rp_pt, "parttree")

# simple plot
plot(rp_pt)

# removing the (recursive) partition borders helps to emphasise overall fit
plot(rp_pt, border = NA)

# customize further by passing extra options to (tiny)plot
plot(
   rp_pt,
   border  = NA,                                     # no partition borders
   pch     = 16,                                     # filled points
   alpha   = 0.6,                                    # point transparency
   grid    = TRUE,                                   # background grid
   palette = "classic",                              # new colour palette
   xlab    = "Topmost vertebra operated on",         # custom x title
   ylab    = "Patient age (months)",                 # custom y title
   main    = "Tree predictions: Kyphosis recurrence" # custom title
)

#
## conditional inference trees from partyit

library("partykit")
ct = ctree(Species ~ Petal.Length + Petal.Width, data = iris)
ct_pt = parttree(ct)
plot(ct_pt, pch = 19, palette = "okabe", main = "ctree predictions: iris species")

## rpart via partykit
rp2 = as.party(rp)
parttree(rp2)

#
## various front-end frameworks are also supported, e.g.

# tidymodels

# install.packages("parsnip")
library(parsnip)

decision_tree() |>
  set_engine("rpart") |>
  set_mode("classification") |>
  fit(Species ~ Petal.Length + Petal.Width, data=iris) |>
  parttree() |>
  plot(main = "This time brought to you via parsnip...")

# mlr3 (NB: use `keep_model = TRUE` for mlr3 learners)

# install.packages("mlr3")
library(mlr3)

task_iris = TaskClassif$new("iris", iris, target = "Species")
task_iris$formula(rhs = "Petal.Length + Petal.Width")
fit_iris = lrn("classif.rpart", keep_model = TRUE) # NB!
fit_iris$train(task_iris)
plot(parttree(fit_iris), main = "... and now mlr3")