Staged Learned Coordinates

This repository contains the learned-coordinate code and paper draft for:

Staged Learned Coordinates for Gradient Boosted Trees on Continuous Tabular Data

The scope is intentionally narrow. The main claim is a robust default recipe for the continuous-tabular slice of boosted-tree problems, centered on complexity-at-target metrics rather than final ensemble size.

Paper

The paper is:

Staged Learned Coordinates for Gradient Boosted Trees on Continuous Tabular Data

Canonical DOI: 10.5281/zenodo.20033698

Recommended public-release artifacts:

• Zenodo preprint record with a DOI
• GitHub repository for code and reproducibility
• PDF artifact: paper/main.pdf
• Source artifact: paper/arxiv_submission_20260404.tar.gz

The repo includes a release checklist at RELEASE_CHECKLIST.md and copy-paste Zenodo metadata at paper/ZENODO_METADATA.md.

License

Code in this repository is released under the MIT License. The paper is released through Zenodo under CC BY 4.0.

Layout

• learned_coordinates/: package code, runners, and tests
• paper/: arXiv-style LaTeX draft plus generated tables and figures

Local experiment outputs are generated when you run the commands below and are ignored by git.

Environment

This repo uses a minimal Python project file at the root.

uv sync

Main commands

Synthetic benchmark:

uv run python -m learned_coordinates.run

Real-data benchmark:

uv run python -m learned_coordinates.run_real

Continuous benchmark comparison:

uv run python -m learned_coordinates.run_continuous_benchmark \
  --results-dir <local-output-dir>/continuous_benchmark_full

Staged comparison:

uv run python -m learned_coordinates.run_continuous_staged_dual_branch \
  --results-dir <local-output-dir>/continuous_staged_dual_branch_full

Branch-union comparison:

uv run python -m learned_coordinates.run_continuous_branch_union \
  --results-dir <local-output-dir>/continuous_branch_union_full

Scalar-gate comparison:

uv run python -m learned_coordinates.run_continuous_regime_gate \
  --results-dir <local-output-dir>/continuous_regime_gate_full

Regenerate paper tables and figures:

uv run python paper/prepare_artifacts.py

Notes

• Paper tables and figures are built from local run outputs; paper/prepare_artifacts.py auto-discovers them and also accepts explicit directory overrides.
• The paper draft can consume a completed held-out run when it exists locally; the current repo now includes a generated held-out comparison based on continuous_heldout_pack_20260404.
• The mixed-data section is secondary and should not be read as a universal tabular claim.