1] My Logic & Dataset Choice

1. Iris (ID 59)

I used this as a control to verify that mlr3oml is pulling data correctly from the OpenML API. It serves as the standard benchmark for the otsk() function.

2. Pen vs. Pencil

I designed this custom task to verify my own understanding of how mlr3 handles mixed feature types. By creating a dataset from scratch with specific attributes (length, ink, material), I could confirm that the task construction logic works exactly as intended before moving on to larger datasets.

2] Implementation

In this section, we load the core mlr3 libraries and execute the task creation logic.

# Loading core libraries
library(mlr3)
library(mlr3oml)
library(data.table)
 
task_iris = as_task(otsk(id = 59))
print(task_iris)
## 
## ── <TaskClassif> (150x5) ───────────────────────────────────────────────────────
## • Target: class
## • Target classes: Iris-setosa (33%), Iris-versicolor (33%), Iris-virginica
## (33%)
## • Properties: multiclass
## • Features (4):
##   • dbl (4): petallength, petalwidth, sepallength, sepalwidth
# Building the dataset from scratch
stationery_data = data.table(
  length_cm     = c(14.5, 15.0, 13.8, 17.5, 18.2, 19.1),
  has_ink       = factor(c("Yes", "Yes", "Yes", "No", "No", "No")),
  has_eraser    = factor(c("No", "No", "No", "Yes", "Yes", "Yes")),
  body_material = factor(c("Plastic", "Plastic", "Metal", "Wood", "Wood", "Wood")),
  label         = factor(c("Pen", "Pen", "Pen", "Pencil", "Pencil", "Pencil"))
)

# Creating the classification task
task_pen_pencil = as_task_classif(stationery_data, target = "label", id = "pen_vs_pencil")

# Printing task details to verify construction
print(task_pen_pencil)
## 
## ── <TaskClassif> (6x5) ─────────────────────────────────────────────────────────
## • Target: label
## • Target classes: Pen (positive class, 50%), Pencil (50%)
## • Properties: twoclass
## • Features (4):
##   • fct (3): body_material, has_eraser, has_ink
##   • dbl (1): length_cm

4] Note: Dataset Selection Rationale

While datasets like TITANIC or CREDIT-G are common choices, I opted for a custom-built PEN VS. PENCIL task.

This decision was made to move beyond standard benchmarks and demonstrate a deeper grasp of the mlr3 workflow. By defining my own features from scratch, I can verify that the conversion from a raw data.table to an mlr3 task functions perfectly with specific data types I controlled, ensuring a more rigorous test of the underlying logic.