afisp-rs

A clean-room Rust + PyO3 implementation of the AFISP workflow: stability analysis to find worst-performing data subsets, then interpretable subgroup phenotyping.

Full documentation: DOCUMENTATION.md

Two Python APIs ship in one package:

Install from PyPI

pip install afisp-rs

Original AFISP compatibility (EBM cross-fit, pandas, plotting, statsmodels):

pip install "afisp-rs[original]"

Demo extras (includes PyTorch surrogate losses):

pip install "afisp-rs[demo]"

Requires Python ≥ 3.9 and numpy. Prebuilt wheels are provided for common Linux, macOS, and Windows platforms. If no wheel matches your platform, pip builds from source and you will need Rust + Python dev headers.

Quick start

Rust-native API (recommended for new code)

import numpy as np
from sklearn.ensemble import RandomForestRegressor
from afisp_rs import WorstSubsetFinder, SubgroupPhenotyper, brier_loss

y_true = np.array([0, 1, 0, 1, 0, 1], dtype=np.uint8)
y_score = np.array([0.2, 0.8, 0.3, 0.7, 0.1, 0.9])
losses = np.asarray(brier_loss(y_true, y_score))

X = np.random.randn(len(losses), 3)

finder = WorstSubsetFinder(subset_fractions=[0.5, 1.0], eps=1e-8, random_state=7)
finder.fit_with_estimator(X, losses, RandomForestRegressor(random_state=7), cv=3)

print("Worst-case risks:", finder.r_hats_)
print("Subset masks:", finder.subset_masks())

subset_labels = finder.subset_masks()[0].astype(np.uint8)
phen = SubgroupPhenotyper(backend="sirus", random_state=7, sirus_num_trees=64)
phen.fit(
    X,
    subset_labels,
    losses,
    y_true=y_true,
    y_score=y_score,
    performance_threshold=0.6,
)
print(phen.summary_table())

Original-style sklearn API

from afisp import WorstSubsetFinder, SubgroupPhenotyper

# Requires: pip install "afisp-rs[original]"
# Uses ExplainableBoostingRegressor by default for conditional-loss cross-fit.

finder = WorstSubsetFinder(subset_fractions=[0.2, 0.4, 1.0], cv=10, eps=1e-5)
finder.fit(X, losses, feature_names=["age", "sex", "score"])
print(finder.R_hats_)
print(finder.confidence_intervals())
idx, effect = finder.find_max_effect_size()

For Rust-backed phenotyping through the compatibility layer, pass y_true, y_score, and performance_threshold to SubgroupPhenotyper.fit().

What is implemented

Stage 1 — Stability analysis

• Fold-aware worst-subset thresholds on out-of-fold conditional loss estimates (mu_hat)
• Influence-function style worst-case risk estimator with analytical 95% CIs
• Subset masks, observed fractions, Cohen's d effect sizes

Stage 2 — Subgroup phenotyping

• backend="sirus" — Rust-native stable rule forest (default)
• backend="conjunction" — deterministic conjunction miner (DecisionList fallback in afisp layer)
• Bootstrap AUROC CIs, Holm–Bonferroni correction, Cohen's d filtering

See DESIGN.md for fidelity notes vs the upstream Python reference.

Rust library (crates.io)

[dependencies]
afisp_rs = { version = "0.1", default-features = false }

Pure Rust usage (no Python):

use afisp_rs::WorstSubsetFinderCore;

let mut finder = WorstSubsetFinderCore::new(vec![0.2, 1.0], 0.0, 0);
finder.fit_from_mu_hat(losses, mu_hat, Some(fold_ids)).unwrap();

Enable the Python extension when building with maturin: --features extension-module.

Development (from source)

git clone https://github.com/na399/afisp-rs.git
cd afisp-rs
python -m pip install maturin pytest numpy scikit-learn
maturin develop --release --features extension-module
pytest -q
cargo test --no-default-features

Full example: examples/demo.py

Publishing

Maintainers: see PUBLISHING.md. Releases are automated on v* tags to PyPI (afisp-rs) and crates.io (afisp_rs).

License

MIT License. See LICENSE.