Random Neighbors

Random Neighbors is a small research repo for random-forest-style clustering / feature selection on high-dimensional data.

It contains:

• RandomNeighbors: repeatedly subsample rows + features, fit a clustering kernel, and keep the feature subsets that produce the best cluster separation (silhouette by default).
• UnsupervisedRFProximity (baseline): an unsupervised random forest trick that learns a proximity matrix from leaf co-occurrence (Breiman-style proximities), then clusters in proximity space.

This makes it easy to run experiments like:

• "Can I find a small feature subset that yields decent clusters?"
• "Does feature bagging beat an unsupervised RF proximity baseline on this dataset?"

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -e .[dev]

Quick start

import numpy as np
from sklearn.cluster import KMeans
from rnc import RandomNeighbors

X = np.random.default_rng(0).normal(size=(1000, 500))

rn = RandomNeighbors(
    select_columns="log2",
    select_rows="sqrt",
    sample_iter=200,
    random_state=0,
)

res = rn.fit(X, clusterer=KMeans(n_clusters=5, n_init="auto", random_state=0))
print(res.best_score, res.best_features)

print(rn.feature_importance(top_k=10))

Experiments

See experiments/.

python -m experiments.run_blobs
python -m experiments.run_iris

Docs

• docs/algorithm.md — what the code is doing
• docs/related_work.md — references and keywords

Status / TODO

• Add more principled feature importance (stability selection / permutation)
• Add benchmark harness + plots
• Add scalable approximations for URF proximity (n can get big quickly)