redumetrics 0.1.0

Metrics to evaluate dimensionality reduction quality.

ReduMetrics

ReduMetrics is a lightweight library to evaluate the quality of dimensionality reductions, independent of the projection method (PCA, t-SNE, UMAP, …).
It provides five complementary metrics:

Metric	Meaning	Range
ULSE	Local neighborhood preservation	[0, 1]
RTA	Random triplet accuracy	[0, 1]
Spearman	Rank correlation of sampled distances	[−1, 1]
k-NCP	k-nearest class preservation	[0, 1]
CDC	Centroid distance correlation	[−1, 1]

Pure functions — NumPy in/out.
Tested with Python 3.9 – 3.12.

Dependencies:
numpy, scipy, scikit-learn

---

Installation

pip install ReduMetrics

---

Usage

from ReduMetrics.metrics.ulse import ulse_score
from ReduMetrics.metrics.rta import rta_score
from ReduMetrics.metrics.spearman import spearman_correlation
from ReduMetrics.metrics.k_ncp import kncp_score
from ReduMetrics.metrics.cdc import cdc_score

# X_high: (m, n) high-dim data, X_low: (m, r) embedding, labels: (m,)
rng = np.random.default_rng(42)
m, n, r = 1000, 50, 2
X_high = rng.normal(size=(m, n))
X_low  = X_high[:, :r]          # toy projection
labels = rng.integers(0, 10, size=m)

# Metrics
ulse = ulse_score(X_high, X_low, k=10)                         # -> [0, 1]
rta  = rta_score (X_high, X_low, T=10000, random_state=0)      # -> [0, 1]
rho  = spearman_correlation(X_high, X_low, P=20000, random_state=0)  # -> [-1, 1]
kncp = kncp_score(X_high, X_low, labels)                       # -> [0, 1]
cdc  = cdc_score (X_high, X_low, labels)                       # -> [-1, 1]

print(ulse, rta, rho, kncp, cdc)