drasc 0.1.0

Density-Ratio Adaptive Spectral Clustering: a scikit-learn-compatible clustering estimator.

DRASC — Density-Ratio Adaptive Spectral Clustering

A small, dependency-light, scikit-learn-compatible clustering estimator.

DRASC builds a sparse k-nearest-neighbour graph with a self-tuning (per-point adaptive) bandwidth, down-weights edges that bridge regions of very different density via a density-ratio term raised to the power gamma, and then runs normalized spectral clustering on the resulting graph.

Installation

pip install drasc

Or from a local checkout:

pip install .

Requires Python ≥ 3.9 and numpy, scipy, scikit-learn (installed automatically).

Quick start

from sklearn.datasets import load_wine
from sklearn.preprocessing import StandardScaler
from drasc import DRASC

X = StandardScaler().fit_transform(load_wine().data)

model = DRASC(n_clusters=3)
labels = model.fit_predict(X)

Because it follows the scikit-learn estimator API, it works with Pipeline, GridSearchCV, and the sklearn.metrics clustering scores:

from sklearn.pipeline import make_pipeline
from sklearn.metrics import adjusted_rand_score

pipe = make_pipeline(StandardScaler(), DRASC(n_clusters=3))
labels = pipe.fit_predict(X)

Parameters

Parameter	Default	Description
n_clusters	—	Number of clusters to form (required, ≥ 2).
n_neighbors	None	Graph neighbours; None → max(10, ceil(2·log2 n)).
gamma	2.0	Density-ratio exponent. 0 disables reweighting.
random_state	0	Seed for the final k-means step.
n_init	30	k-means initializations on the embedding.
n_jobs	-1	Parallel jobs for the neighbour search.

Fitted attributes

• labels_ — cluster label per sample.
• embedding_ — row-normalized spectral embedding (n_samples, n_clusters).
• eigengap_ — relative eigengap at the n_clusters boundary, useful for selecting n_clusters / gamma / n_neighbors.

Tips

• Standardize your features first (e.g. StandardScaler). DRASC uses Euclidean distances, so feature scales matter.
• For high-dimensional data, reduce with PCA before clustering.
• Sweep eigengap_ over candidate settings to pick hyperparameters without labels.

License

MIT