nacvi 0.1.0

Cluster Validity Indices for Noise-Aware Clusterings (e.g. DBSCAN)

Noise-Aware Cluster Validity Indices (NACVI)

This repository contains a Python implementation of internal cluster validity indices specifically designed for noise-aware Clusterings (e.g. DBSCAN). The validity indices presented here explicitly consider unassigned data points (noise), which makes them particularly suitable for realistic, unsupervised settings.

This is based on the scientific publication:
Lea Eileen Brauner, Frank Höppner, Frank Klawonn
Cluster Validity for Noise-Aware Clusterings, Intelligent Data Analysis Journal, IOS Press (2025)

---

Content

You can find the implementations of the following NACVIs:

• sil+: noise-aware Silhouette Coefficient
• dbi+: noise-aware Davies-Bouldin Index
• gD33+: noise-aware Dunn-Index-Variant
• sf+: noise-aware Score Function
• grid+: grid-based noise-validity index
• nr+: neighbourhood-based Noise-validity index

---

Motivation

Conventional validity measures treat all data points as belonging to a cluster, even if noise is explicitly labelled in DBSCAN, for example. This leads to distorted evaluations.

This package:

• takes noise into account correctly,
• enables a separate evaluation of the cluster structure and the noise delimitation,
• offers an integrated metric for both with the B+ score.

---

Installation

pip install nacvi

Usage

In examples/usage_miniexample.py you can find a minimal example for the usage with numpy arrays as inputs.

In examples/usage_example.ipynb you can find a comprehensive example with:

• data generation,
• execution of the DBSCAN clustering algorithm,
• visualisation,
• calculation of the NACVIs