Visualize fitted oriented hyperrectangles

Creates pairwise scatter plots showing fitted oriented (rotated) hyperrectangles overlaid on hyperparameter configurations. Uses convex hull projections to visualize the rotated hyperrectangle in 2D pairwise plots.

Usage

# S3 method for class 'LearnerSubspacePolygon'
autoplot(
  object,
  select = "all",
  force = FALSE,
  wrap = TRUE,
  size_top = 0.7,
  size_all = 0.5,
  ...
)

Arguments

object

A trained LearnerSubspacePolygon object with fitted subspace

select

Character vector of hyperparameter names to plot, or "all" (default) to plot all hyperparameters

force

Logical indicating whether to skip the confirmation prompt when plotting many hyperparameters. Default: FALSE

wrap

Logical indicating whether to combine plots using patchwork::wrap_plots(). Default: TRUE

size_top

Numeric point size for top-performing configurations (orange crosses). Default: 0.7

size_all

Numeric point size for all data points (gray background). Default: 0.5

...

Additional arguments passed to patchwork::wrap_plots() (only used when wrap = TRUE)

Value

If wrap = TRUE: A single patchwork object combining all plots.

If wrap = FALSE and no categorical hyperparameters: A list of ggplot objects, one per hyperparameter pair.

If wrap = FALSE and categorical hyperparameters present: A named list where each element is a list of ggplot objects for that categorical level.

Details

Plot Structure:

Each plot shows:

• Gray points: All configurations in the dataset (low alpha)

• Orange crosses: Top-performing configurations used for fitting

• Blue polygon: Projection of the oriented hyperrectangle onto 2D plane

• Blue diamond: Center of the hyperrectangle

For \(p\) selected hyperparameters, creates \(\binom{p}{2}\) pairwise plots.

Visualization Method:

The oriented hyperrectangle is defined by transformation matrix \(A\) and translation \(b\). For visualization:

1. Computes all \(2^p\) vertices of the unit hypercube \([-1,1]^p\)

2. Transforms vertices to original space: \(x = A^{-1}y + c\) where \(c = -A^{-1}b\)

3. Projects vertices onto each 2D hyperparameter pair

4. Computes convex hull of projected vertices for visualization

Univariate Case:

When only one hyperparameter is selected, displays a histogram with vertical lines marking the hyperrectangle boundaries and rug plot for top configurations.

Categorical Hyperparameters:

When the task includes a categorical hyperparameter, separate plots are created for each categorical level, showing the corresponding fitted hyperrectangle.

Interactive Prompt:

When plotting more than 3 hyperparameters with wrapping enabled, the function prompts for confirmation due to potential readability issues. Use force = TRUE to bypass this prompt.

Dependencies:

Requires ggplot2, grDevices, and scales. If wrap = TRUE, also requires patchwork.

See also

LearnerSubspacePolygon for the learner class. coef.LearnerSubspacePolygon for extracting fitted parameters. autoplot.LearnerSubspaceBox for axis-aligned visualization.

Examples

if (FALSE) { # \dontrun{
# Train learner
task <- TaskSubspace$new(data, target_measure = "auc",
                         hps = c("learning_rate", "max_depth"))
learner <- LearnerSubspacePolygon$new(task)
learner$train(q_val = 0.9, lambda = 0.1)

# Plot all hyperparameters (wrapped)
autoplot(learner)

# Plot specific hyperparameters
autoplot(learner, select = c("learning_rate", "max_depth"))

# Get individual plots without wrapping
plots <- autoplot(learner, wrap = FALSE)
plots[[1]]  # First pairwise plot

# Customize wrapping layout
autoplot(learner, ncol = 2, guides = "collect")

# With categorical hyperparameters
task <- TaskSubspace$new(data, target_measure = "auc",
                         hps = c("learning_rate", "max_depth"),
                         cat_hps = "optimizer")
learner <- LearnerSubspacePolygon$new(task)
learner$train(q_val = 0.9)
autoplot(learner)  # Separate plots per optimizer
} # }