Purity<T>

Models & Types · AiDotNet.Clustering.Evaluation

Purity metric for evaluating clustering against ground truth labels.

Inheritance: ExternalClusterMetricBase<T> → Purity<T>

For Beginners

Purity asks “How pure are the clusters?”

For each cluster:

  • Find the most common true class
  • Count how many points belong to that class

Purity = Total correctly assigned / Total points

Example:

  • Cluster 1: 30 cats, 10 dogs → 30 correct (majority is cats)
  • Cluster 2: 20 cats, 40 dogs → 40 correct (majority is dogs)
  • Purity = (30 + 40) / 100 = 0.70

Higher purity = Better clustering! But beware: Purity increases as k increases (k=n gives purity=1).

How It Works

Purity measures the fraction of correctly assigned points, where “correct” means each cluster is assigned to its majority class. Values range from 1/k (random) to 1 (perfect).

Purity = (1/n) * sum over all clusters k of max_j |c_k ∩ t_j| Where c_k is cluster k and t_j is true class j.

Constructors

Constructor Summary
Purity() Initializes a new Purity instance.

Methods

Method Summary
Dictionary<int, double> ComputePerCluster(Vector<T> trueLabels, Vector<T> predictedLabels) Computes purity per cluster.
double Compute(Vector<T> trueLabels, Vector<T> predictedLabels)