Notes
🔍

Linkage Criterion

When hierarchical clustering builds groups, it has to answer a very practical question: if two clusters each contain several points, how do we measure the distance between the clusters themselves? That rule is the linkage criterion, and it strongly shapes the clusters you end up seeing.

What it means

In agglomerative hierarchical clustering, the algorithm starts with every data point as its own cluster and repeatedly merges the two closest clusters. The key detail is how “closest” is defined. A linkage criterion is the formula used to compute inter-cluster distance from the point-to-point distances inside and across clusters.

  • Single linkage: distance between the nearest pair of points across two clusters.
  • Complete linkage: distance between the farthest pair of points.
  • Average linkage: average of all pairwise distances across the two clusters.
  • Ward linkage: merges clusters that cause the smallest increase in within-cluster variance.

Why it matters

This choice is not a minor setting. It changes the geometry of the result. Single linkage can create long “chains,” which helps connect irregular shapes but can also merge through noise. Complete linkage prefers tight, compact clusters. Average linkage is a middle ground. Ward linkage usually produces balanced, spherical-looking groups and is widely used for customer segmentation and document grouping when Euclidean structure makes sense. If you ignore linkage choice, your dendrogram can suggest a very different story about the same data.

In practice

In tools like scikit-learn, you see this directly in AgglomerativeClustering(linkage=...), and in SciPy’s scipy.cluster.hierarchy.linkage. For transaction patterns, a poor linkage can merge unrelated customers too early. For gene-expression or topic-discovery data, it can hide meaningful subgroups or exaggerate weak ones. The linkage criterion is the rule that decides what “belongs together” at every merge step, so it is one of the most important design choices in hierarchical clustering.

Linkage Criterion is the rule in hierarchical clustering that defines the distance between two clusters, determining which clusters merge at each step. Common choices include single, complete, average, and Ward linkage, each producing different cluster shapes and dendrogram structures. It matters because the linkage criterion directly controls the clustering outcome: a poor choice can merge dissimilar groups or fragment coherent ones, while a suitable one preserves meaningful structure in unlabeled data.

Imagine you’re grouping towns on a map. Do you decide two groups belong together because their closest houses are near each other, because their farthest houses are still fairly close, or because the average distance between all houses is small? That rule is the linkage criterion.

In AI clustering, a linkage criterion is the rule for deciding when two groups of data should be treated as one bigger group. It matters because different rules create different kinds of clusters. One rule may form long, chain-like groups, while another prefers tight, compact ones. So this simple choice strongly shapes how the AI “sees” natural groupings in unlabeled data.