Bankai Random Forest

Bankai is a scikit-learn-compatible random forest classifier with a native Rust backend. The project is in alpha; APIs and serialized models may change. The roadmap and compatibility details are in docs/ROADMAP.md and docs/SKLEARN_COMPATIBILITY.md.

Install from source

Python 3.11 or newer and a Rust toolchain are required to build the extension. From the repository root:

uv sync
uv run maturin develop --release

The --release flag builds the optimized native extension. Then install the package in your Python environment using the workflow above and import it as shown below.

Quick start

BankaiRandomForestClassifier follows the familiar scikit-learn estimator API:

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

from bankai_random_forest import BankaiRandomForestClassifier

X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42, stratify=y
)

model = BankaiRandomForestClassifier(
    n_estimators=100,
    random_state=42,
    max_bins=64,  # omit this or use None for exact split search
)
model.fit(X_train, y_train)

predictions = model.predict(X_test)
probabilities = model.predict_proba(X_test)
print(f"Accuracy: {accuracy_score(y_test, predictions):.3f}")
print("Feature importances:", model.feature_importances_)

The classifier works with scikit-learn utilities such as Pipeline, cross_val_score, and GridSearchCV. Features should be numeric; encode categorical columns with a transformer such as OneHotEncoder in a pipeline.

Common options

• n_estimators, max_depth, min_samples_split, min_samples_leaf, max_features, bootstrap, max_samples, class_weight, and random_state control forest fitting.
• max_bins=None uses exact split search. An integer from 2 to 255 enables histogram split search; for example, max_bins=64.
• ccp_alpha prunes trees. monotonic_cst accepts one -1, 0, or 1 per feature and is supported for binary classification.
• oob_score=True enables out-of-bag estimates when bootstrap=True.
• importance_type selects "gain" (default), "split", or "permutation". Permutation importance requires bootstrap=True.
• Dense arrays and SciPy CSR/CSC matrices are supported, including NaN feature values. Missing values are routed by the fitted trees; categorical values are not automatically encoded.

Multioutput and multilabel targets are supported. Bankai trains one native forest per output, whereas scikit-learn shares tree structures across outputs; therefore trees and probabilities need not match exactly. For multioutput models, estimators_, apply, decision_path, and TreeSHAP inspect the first output forest. Monotonic constraints currently apply only to binary single- output classification.

TreeSHAP

Install SHAP in the same environment (python -m pip install shap), then pass the fitted classifier directly to TreeExplainer:

import shap

explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

This integration supports single-output classifiers. See docs/SKLEARN_COMPATIBILITY.md for the multioutput inspection limitation and compatibility status.

Save and restore a fitted model

import joblib

joblib.dump(model, "classifier.joblib")
restored_model = joblib.load("classifier.joblib")
predictions = restored_model.predict(X_test)

Bankai's joblib state includes the training arrays and rebuilds the native forest while loading, so loading has reconstruction cost and artifacts may be large. Keep Bankai, Python, and dependency versions consistent when restoring models. Only load joblib files from trusted sources. Details are in docs/JOBLIB_COMPATIBILITY.md.

Development and tests

uv run pytest -q
uv run cargo test --workspace

Install benchmark-only dependencies with uv sync --group benchmark.

Optimized cp311-abi3 wheels were built and tested locally on Python 3.11, 3.12, and 3.13. The full Python suite passed (147 tests) against each installed wheel. Release signing remains future work.

The project is licensed under GPL-3.0-or-later; see COPYING.