Course contentsShow
← Machine Learning and Deep Learning
Lesson 357 of 3,5388. Unsupervised Learning: ClusteringFree lesson

Linkage Criteria: Single, Complete, and Average

Comparing linkage methods that define distance between clusters—single (minimum), complete (maximum), average.

Linkage Criteria: Single, Complete, and Average

What you'll learn: How hierarchical clustering decides which clusters to merge by measuring distances in three different ways.

The Core Problem

When hierarchical clustering builds a dendrogram, it repeatedly merges the two "closest" clusters. But here's the question: how do you measure distance between groups of points? A cluster contains multiple data points, not just one. The linkage criterion determines how we calculate this cluster-to-cluster distance.

The Three Main Linkage Methods

Single Linkage (Minimum)

Measures distance as the shortest distance between any two points from different clusters—like finding the two closest neighbors across a border.

Analogy: Two countries are "close" if their nearest border towns are close, even if their capitals are far apart.

Behavior: Tends to create long, chain-like clusters. Sensitive to noise and outliers because one stray point can bridge distant groups.

Complete Linkage (Maximum)

Measures distance as the longest distance between any two points from different clusters—the farthest pair defines separation.

Analogy: Two countries are only "close" if even their most distant points aren't too far apart.

Behavior: Prefers compact, spherical clusters. More robust to outliers but may split large natural clusters.

Average Linkage

Measures distance as the average distance between all pairs of points from different clusters.

Analogy: Two countries' closeness is determined by the average distance between all possible town pairs.

Behavior: Balanced approach between single and complete. Often produces well-separated, reasonably compact clusters.

Choosing Your Linkage

  • Single: When you expect elongated, winding cluster shapes
  • Complete: When clusters should be tight and well-separated
  • Average: General-purpose, balanced option for most cases

Key Takeaway: Linkage criteria define how to measure distance between clusters; single uses minimum distance (chain-prone), complete uses maximum distance (compact clusters), and average uses mean distance (balanced compromise).