hyperDT

Fast hyperboloid decision tree algorithms

This repository contains code for the paper Fast Hyperboloid Decision Tree Algorithms (ICLR 2024), which you can find at one of these links:

• OpenReview
• ArXiv

To install, run this from your repo directory:

git clone https://github.com/pchlenski/hyperdt
cd hyperdt
pip install -e .

All dependencies are managed through pyproject.toml. For development, you can install the package with all development dependencies:

pip install -e ".[dev]"

For testing, you can install the test dependencies:

pip install -e ".[test]"

Additionally, hyperDT is available on PyPI. It can be pip installed as follows:

pip install hyperdt

HyperDT supports optional dependencies:

# Install with XGBoost support
pip install hyperdt[xgboost]

# Install with legacy implementation support (requires geomstats)
pip install hyperdt[legacy]

# Install all optional dependencies
pip install hyperdt[all]

A basic tutorial demonstrating key HyperDT functionality is available in notebooks/tutorial.ipynb.

The hyperDT package is structured as follows:

• hyperdt
  • _base.py: Base classes for hyperbolic decision trees
  • toy_data.py: Utilities for generating synthetic hyperbolic datasets
  • tree.py: HyperbolicDecisionTreeClassifier and HyperbolicDecisionTreeRegressor
  • ensemble.py: HyperbolicRandomForestClassifier and HyperbolicRandomForestRegressor
  • oblique.py: Oblique decision trees (requires pip install hyperdt[oblique])
  • xgboost.py: XGBoost integration (requires pip install hyperdt[xgboost])
  • legacy/: Original implementation (requires pip install hyperdt[legacy])
    • dataloaders: Functions for loading benchmarking data into HoroRF
    • conversions: Convert between hyperboloid, Poincare, and Beltrami-Klein models
    • ensemble: Legacy HyperbolicRandomForestClassifier and HyperbolicRandomForestRegressor
    • hyperbolic_trig: Angular processing and midpoint calculations in the hyperboloid model
    • tree: Legacy HyperbolicDecisionTreeClassifier and HyperbolicDecisionTreeRegressor, plus base classes
    • visualization: Code to visualize decision boundaries on the Poincare disk
• tests/: Test files for verifying functionality
  • test_typing.py: Type annotation verification tests
  • test_model_types.py: Tests for classifier and regressor functionality
  • test_toy_data.py: Tests for data generation utilities
  • test_equivalence.py: Tests comparing the new implementation to the legacy code
  • test_oblique_models.py: Tests verifying oblique decision trees can train
  • test_midpoint_override.py: Verifies models are editable by zeroing out midpoints

The package has a modular design with optional dependencies:

• Core functionality only requires numpy, scikit-learn, scipy, and matplotlib
• XGBoost backend requires the xgboost package (pip install hyperdt[xgboost])
• Legacy implementation requires geomstats (pip install hyperdt[legacy])

All figures and tables in the paper were generated using a combination of Python scripts and Jupyter notebooks. The notebooks used in development were filtered down to only those that remained relevant to the final paper and moved to the notebooks/archive directory. The notebooks directory contains a tutorial and symbolic links to notebooks of particular relevance to a figure, table, or section of a paper, named according to the section they reproduce.

benchmarks/hororf_benchmarks.py runs the benchmarks contributing to Tables 1, 5, and 6, and benchmarks/scaling_benchmarks.py runs the benchmarks contributing to Figures 6 and 7.

All relevant datasets, plus benchmarking code outputs, can be found on Google Drive.

To cite HyperDT, please use the following:

@inproceedings{
    chlenski2024fast,
    title={Fast Hyperboloid Decision Tree Algorithms},
    author={Philippe Chlenski and Ethan Turok and Antonio Khalil Moretti and Itsik Pe'er},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=TTonmgTT9X}
}