Purity-Threshold-Pruning 0.1.0

A pruning algorithm for the scikit-learn RandomForestClassifier which minimizes model complexity while considering a user-specified tolerance for accuracy loss.

Purity-Threshold-Pruning

About

Purity Threshold Pruning is a bagging‑based tree‑ensemble pruning algorithm for scikit‑learn’s RandomForestClassifier that minimizes model complexity under a user‑specified accuracy‑loss tolerance.

The main advantages are:

• Explicitly consider node purity (i.e., class distribution), preventing the pruning of low purity nodes that may cause catastrophic accuracy degradation.
• Iteratively search for the minimum purity threshold that satisfies the tolerance, resulting in the smallest possible model that satisfies the desired performance.
• Systematical pruning prevents the model from overgrowing or being over-pruned.

Installation

From PyPI (recommended)

pip install purity-threshold-pruning

From GitHub

git clone https://github.com/Datou0718/Purity-Threshold-Pruning
cd Purity-Threshold-Pruning
pip install .

Quick Start

from PurityThresholdPruning import PurityThresholdPruning
from sklearn.ensemble import RandomForestClassifier

# train a Random Forest
rf = RandomForestClassifier(...)
rf.fit(X_train, y_train)

# Pruned the model with 5% tolerable accuracy loss
tolerance = 0.05 
rf = PurityThresholdPruning(rf, tolerance, X_train, y_train)

API

PurityThresholdPruning(forest, tolerance, X_train, y_train)

• forest: a fitted RandomForestClassifier with oob_score=True.
• tolerance: Maximum tolerable accuracy drop in OOB estimation (e.g., 0.05 for 5%).
• X, y: Training data (np.ndarray) used to compute OOB score. Returns the RandomForestClassifier pruned with the minimum purity threshold.