Getting Started with classbound

What is classbound?

classbound is an R package for exploring and comparing classification decision boundaries. Given a fitted classifier and a dataset, it answers a simple question: where in the feature space does the model change its prediction?

The package supports:

Installation

devtools::install_github("natydasilva/classbound")

Quick start: one-step boundary plot

The classbound() wrapper fits a model, computes the decision boundary, and plots it in a single call.

library(classbound)
library(palmerpenguins)

penguins <- na.omit(penguins[, c("species", "bill_length_mm", "bill_depth_mm")])

classbound(
  data       = penguins,
  formula    = species ~ bill_length_mm + bill_depth_mm,
  classifier = rpart::rpart
)

The colored regions show the predicted class for each point in the feature space. Training observations are overlaid as points colored by their true class.

The modular API

For more control, use the three-step pipeline:

# Step 1: Fit the model
model <- fit_model(penguins, species ~ bill_length_mm + bill_depth_mm, rpart::rpart)

# Step 2: Compute the boundary grid
model <- boundary_compute(
  model,
  feature_range = list(bill_length_mm = c(30, 60), bill_depth_mm = c(10, 25)),
  resolution    = 80
)

# Step 3: Plot
plot_boundary(
  model,
  obs_data   = penguins,
  x_col      = "bill_length_mm",
  y_col      = "bill_depth_mm",
  true_label = "species"
)

boundary_compute() returns the same model object with $boundary_data populated. This means you can reuse the same fitted model with different visualizations, zoom levels, or color settings without refitting.

Using different classifiers

classbound works with any classifier whose predict() method returns a vector of class labels. Most classifiers work automatically:

# SVM (returns class labels natively; no extra work needed)
classbound(penguins, species ~ bill_length_mm + bill_depth_mm, e1071::svm)

# Random forest (matrix interface: randomForest expects x and y separately)
classbound(penguins, species ~ bill_length_mm + bill_depth_mm,
  randomForest::randomForest,
  interface = "matrix"
)

For classifiers whose predict() returns a list or other complex object, use the predfun argument to extract the class labels:

# MASS::qda returns a list, so extract $class manually
classbound(
  penguins,
  species ~ bill_length_mm + bill_depth_mm,
  MASS::qda,
  predfun = function(model, newdata, ...) predict(model, newdata, ...)$class
)

Probability surface

Classifiers that return class probabilities (e.g., rpart, randomForest) enable a gradient visualization where decision regions are shaded by model confidence:

model <- fit_model(penguins, species ~ bill_length_mm + bill_depth_mm, rpart::rpart)
model <- boundary_compute(model)

plot_boundary(
  model,
  obs_data      = penguins,
  x_col         = "bill_length_mm",
  y_col         = "bill_depth_mm",
  true_label    = "species",
  show_gradient = TRUE
)

Deep colors indicate high model confidence; faded colors near boundaries indicate uncertainty.

Classifiers that do not provide probabilities (e.g., standard SVMs, PPtree) always show flat solid regions regardless of show_gradient = TRUE.

Interactive exploration

For interactive exploration without writing code, launch the built-in Shiny application:

# Launch with a dataset pre-loaded
explorapp(data = penguins, target_col = "species")

# Or launch empty and simulate data interactively
explorapp()

See the vignette("explorapp-guide") for a full walkthrough of the interactive features.

What’s next?