MerQuR+Statistics Glossary

Statistics Glossary → Linkage method (Ward, complete, average, single)

Linkage method (Ward, complete, average, single)concept

In hierarchical clustering, the rule that determines how the distance between two clusters is computed from the distances between their members.

Agglomerative hierarchical clustering starts with every observation as its own cluster and at each step merges the two closest clusters. What counts as 'closest' depends on the linkage method. In single linkage the distance between two clusters is the distance between their two nearest members; in complete linkage, between their two farthest members; in average linkage (UPGMA), the average distance over all pairs of members. Ward's method merges the two clusters whose union produces the smallest increase in the within-cluster sum of squares. The centroid and median methods use the distance between cluster centres.

The choice of method can change the result markedly. Single linkage tends to produce long, chain-like clusters and is sensitive to noise; complete linkage forms compact clusters of roughly equal diameter and is affected by outliers; average linkage strikes a balance between the two. Ward's method yields spherical clusters of roughly equal size and is designed for use with Euclidean distance. The centroid and median methods can produce 'inversions' in the dendrogram (merge heights that decrease). There is no single correct method; the choice should follow the structure of the data and the expected shape of the clusters, and the stability of the results should be checked with different methods.

An everyday example

A teacher wants to divide a class into friendship groups. Under single linkage, two groups merge if any pupil in one is a close friend of any pupil in the other; this can link the whole class into one group through a chain of 'my friend's friend's friend'. Under complete linkage, two groups merge only if all their members are close enough to one another, giving tighter, smaller groups.

In MerQur

It is chosen with the 'Linkage method (Ward / Complete / Average / Single / Centroid / Median / McQuitty)' parameter in Hierarchical Clustering; in Variable Clustering, one of the 'Algorithm' options is average linkage with the 1 - |r| distance.

Analyses where this concept appears

Variable Clustering (VARCLUS)Hierarchical Clustering

Related concepts

Dendrogram and cophenetic correlationDistance measure (Euclidean, Manhattan, Gower, great-circle)