A systematic, AI-assisted conversion of the R package rpart to Python.
This repository documents and contains the complete conversion of rpart (version 4.1.27) — the reference implementation of the CART (Classification and Regression Trees) methodology of Breiman, Friedman, Olshen, and Stone (1984) — into an installable Python package, r2py_rpart. The full methodology, findings, and results are documented in the accompanying technical report, docs/report/main.pdf.
rpart is one of the most widely used machine learning libraries in the R ecosystem: a recommended package shipped with every R installation, providing decision-tree fitting for continuous, categorical, count, and survival responses. Despite the centrality of decision trees to modern machine learning, no faithful Python port of rpart existed that preserved its full feature set — surrogate splits for missing-data handling, four built-in splitting methods, a user-defined split callback mechanism, cost-complexity pruning, and the complete cross-validation and visualization infrastructure.
Rather than reimplementing the CART algorithm from scratch, this project reuses rpart's original, validated C source code directly. The C sources are compiled as a standalone shared library — behind a layer of fake R C API headers that removes the runtime dependency on libR.so — and called from Python via cffi. The R source layer, all 47 functions across 36 files, is separately converted to Python function by function, in a dependency-aware order, guided by 172 machine-generated language-construct conversion guides. This strategy confines the entire correctness burden to the narrow, five-entry-point R-to-C interface, and produces a self-contained package that requires neither an R runtime nor any external bridging layer such as rpy2 or reticulate at run time.
The resulting package's test suite comprises 846 tests spanning its full public interface, of which 844 pass and 2 are permanently marked xfail, documenting a single, structurally unfixable formatting divergence from R (R's dput()-based unevaluated-call-expression printing).
| Path | Contents |
|---|---|
rpart/ |
Unmodified R package source (rpart version 4.1.27), as obtained from CRAN; the canonical reference against which every phase of the conversion is checked. |
r2py_rpart/ |
The installable Python package. Managed as a git subtree linked to github.com/r2py-project/r2py_rpart, so it can be versioned and published to PyPI independently of this repository. |
c_refactor_analysis/ |
C dependency-graph analysis artifacts (Phase 1). |
r_extern_analysis/ |
R external item extraction CSVs, fake-header implementation guides, and fake-header conversion guides (Phases 2–4). |
structural_analysis/ |
R structural dependency analysis JSONs and dependency-graph artifacts (Phase 6). |
language_dependency_analysis/ |
Per-file language-dependency CSVs and 172 R-to-Python conversion guides (Phase 8). |
conversion_results/ |
Per-function JSON translation artifacts (Phase 9.1). |
docs/ |
Architecture notes, planning documents, per-phase summaries (daily_summaries/), and the technical report (report/). |
.claude/agents/, .claude/commands/ |
Sub-agent and skill (slash command) specifications for the AI-assisted development workflow used throughout this project. |
git_pull.sh, git_push.sh |
Cluster batch scripts encapsulating the git subtree synchronization commands that keep the python-rpart and r2py_rpart remotes in sync. |
install_environments.sh |
Cluster batch script provisioning the r-to-python conda environment used throughout every phase of the conversion. |
Further detail on the conversion methodology, the thirteen-phase project timeline, and the internal layout of the r2py_rpart package itself is given in the technical report.
pip install r2py_rpartSee r2py_rpart/README.md for installation details, a quick-start example, and the full public API.
docs/report/main.pdf— the technical report: full methodology, phase-by-phase findings, and discussion.docs/daily_summaries/— per-session activity summaries, the primary source material for the technical report.docs/architecture-analysis.md— architectural notes on therpartpackage.
This repository bundles the unmodified rpart R package source (rpart/), which is distributed by CRAN under the GNU General Public License, version 2 or 3 (GPL-2 | GPL-3). The r2py_rpart Python package and all original project artifacts in this repository are, consistently, distributed under the GNU General Public License, version 2 or later (GPL-2.0-or-later). See LICENSE for the full text and r2py_rpart/NOTICE for attribution to the original rpart authors.
The full methodology and findings are documented in the technical report "Converting the R Package rpart to Python: A Technical Report" (2026), included at docs/report/main.pdf.
This project was developed using Claude Code (Anthropic) as an AI-assisted development environment, and used computational resources provided by the HPC facility of the University of Notre Dame.
- Yufei Cai (ycai9@nd.edu)
- Jun Li (jun.li@nd.edu)
University of Notre Dame