GeoRF — Geometry-Aware Random Forest

GeoRF replaces bootstrap resampling with HVRT-powered generative diversity. Each tree trains on a completely unique synthetic dataset drawn from learned per-partition kernel density estimates. No tree ever sees a real sample. No two trees share a single training point.

Why?

Bootstrap bagging has a diversity ceiling. With n = 250 samples, each bootstrap draw contains ≈158 unique samples. GeoRF removes this ceiling: 100 trees × 500 samples = 50 000 unique synthetic training points.

See benchmark/results/ for full benchmark results comparing GeoRF against Random Forest, Gradient Boosting, XGBoost, LightGBM, and MLP (sklearn + PyTorch) on standard datasets. Run the benchmarks yourself:

cd benchmark
pip install -r requirements.txt
python run_classification.py
python run_regression.py

Install

pip install -e .          # editable
# or
pip install georf

Requirements: Python ≥ 3.10, hvrt >= 2.3.0, scikit-learn, numpy, joblib

Quick start

from georf import GeoRFClassifier, GeoRFRegressor

# Classification
clf = GeoRFClassifier(n_estimators=100, n_samples_per_tree=500, n_jobs=-1)
clf.fit(X_train, y_train)
y_pred  = clf.predict(X_test)
y_proba = clf.predict_proba(X_test)   # (n_samples, n_classes)

# Regression
reg = GeoRFRegressor(n_estimators=100, n_samples_per_tree=500, n_jobs=-1)
reg.fit(X_train, y_train)
y_pred = reg.predict(X_test)

# Interpretability
clf.feature_importances(feature_names=cols)  # dict, sorted descending
clf.tree_quality_scores()                    # per-tree AUC array
clf.diversity_score()                        # float: pairwise disagreement rate
clf.provenance()                             # dataset / expansion metadata

Parameters

Running tests

pip install pytest
pytest tests/
# exclude slow timing test:
pytest tests/ -m "not slow"

License

AGPL-3.0-or-later