bisection-regressor 0.2.0

A scikit-learn-compatible regressor that predicts continuous targets via a cascade of binary classifiers bisecting the target's value range.

BisectionRegressor

A scikit-learn-compatible regressor that predicts continuous targets by recursively bisecting the target's value range, using a binary classifier at each node to decide which half a sample falls into. Predictions are formed by softly blending both branches at every node, weighted by the classifier's confidence (via the law of total expectation), rather than committing to a single hard path.

Works as a drop-in scikit-learn estimator: Pipeline, GridSearchCV, cross_val_score, clone, sample_weight are all supported.

Install

From PyPI:

pip install bisection-regressor

From source, for development:

git clone https://github.com/vihan015/bisection-regressor.git
cd bisection-regressor
pip install -e .

To also run the benchmarks (needs pandas):

pip install -e ".[benchmarks]"

Quick start

from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from bisection_regressor import BisectionRegressor

X, y = load_diabetes(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25, random_state=0)

reg = BisectionRegressor(max_depth=6)
reg.fit(X_train, y_train)
print(reg.score(X_test, y_test))   # R^2

Documentation

• docs/hyperparameters.md: every hyperparameter explained, plus usage examples covering Pipeline, GridSearchCV (including BisectionForestRegressor's nested params), cross-validation, sample weights, and tree inspection.
• docs/math_deepdive.md: a from-scratch walkthrough of exactly what fit() and predict() compute internally, and every mathematical concept involved (medians, entropy, the law of total expectation, empirical Bayes shrinkage, probability calibration, bagging), with a fully worked, verified numeric example.

Results

See paper/report.md for the full write-up: how the method was diagnosed and improved from an initial naive design, the mathematical justification for each fix, and 5-fold cross-validated results on three real UCI datasets (Wine Quality, Concrete Compressive Strength, Auto MPG) against standard baselines (Linear Regression, Random Forest, Gradient Boosting, and so on), reported honestly, including where the method does not win.

Reproduce the benchmark yourself: see benchmarks/.

Status

This is a research/experimental estimator, not a production-hardened replacement for gradient boosting or random forests. It's a specific, interpretable structural alternative. See paper/report.md for exactly where it's competitive and where it isn't.

License

MIT. See LICENSE.