spacefinder

Overview

spacefinder identifies promising subspaces within hyperparameter spaces by fitting geometric regions (hyperrectangles and ellipsoids) to high-performing configurations from hyperparameter tuning benchmarks.

Key features:

• Three learner types: Box (axis-aligned), Polygon (oriented), and Ellipsoid
• Handles categorical hyperparameters by fitting separate subspaces per level
• Beta density estimation for probabilistic sampling (augment())
• Visualization tools for fitted subspaces (autoplot())
• Works with multi-task benchmark data

Installation

Install the development version from GitHub:

# install.packages("pak")
pak::pak("NikoGerman/spacefinder")

Quick Start

library(spacefinder)

# Load example data
data(benchmark_data)

# Create task
task <- TaskSubspace$new(
  data = benchmark_data,
  target_measure = "auc",
  hps = c("learning_rate", "max_depth")
)

# Fit axis-aligned hyperrectangle
learner <- LearnerSubspaceBox$new(task)
learner$train(q_val = 0.9)

# View fitted bounds
coef(learner)
#>    hyperparameter         min         max
#>            <char>       <num>       <num>
#> 1:  learning_rate 0.001746421  0.00554191
#> 2:      max_depth 3.000000000 15.00000000

# Visualize fitted subspace
ggplot2::autoplot(learner)

Documentation

Learn more about spacefinder:

• vignette("getting-started", package = "spacefinder") - Basic workflow and concepts
• vignette("learner-comparison", package = "spacefinder") - Comparing Box, Polygon, and Ellipsoid learners
  
• vignette("categorical-hyperparameters", package = "spacefinder") - Working with categorical variables
• vignette("density-estimation", package = "spacefinder") - Probabilistic modeling with augment()

Or browse online at https://nikogerman.github.io/spacefinder/

Citation

If you use spacefinder in your research, please cite:

@Manual{,
  title = {spacefinder: Subspace Learning for Hyperparameter Optimization},
  author = {Nikolai German},
  year = {2025},
  note = {R package version 0.2.0},
  url = {https://github.com/NikoGerman/spacefinder},
}

License

Apache License v2.0