A linkage criterion defines the distance between clusters during hierarchical clustering. Each criterion changes which observations are merged as the algorithm builds the hierarchy. section 2.4.2 section 2.10
Common choices include:
- Single linkage uses the greatest proximity between clusters by considering their closest neighboring objects. section 2.4.2
- Complete linkage uses the greatest distance between two objects in different clusters. section 2.4.2
- Group average linkage averages distances across all possible object pairs in two clusters. section 2.4.2
- Ward linkage minimizes the variability of clusters merged into one. section 2.4.2
Comparing linkage methods in Python
Create hierarchical clusterings from standardized data with SciPy using average, Ward, and complete linkage, then plot each dendrogram. Select a distance threshold, use fcluster with criterion="distance", and plot the resulting labels. Test several distance values because both the number and shape of clusters can change. section 2.4.5
Figure 2.18 Single linkage is for closest neighbors (left); complete linkage is for farthest neighbors (center); and group average is for the average of the distance between clusters (right).
Compare the geometric intuition behind single, complete, and group average linkage. section 2.4.2
