How is it different from k-means?

K-means needs the group count in advance; agglomerative clustering builds the full hierarchy and lets you choose the cut after seeing it.

What is a dendrogram?

The tree diagram of merges, where join heights show how similar merged clusters were.

What data size suits it?

Thousands of items comfortably; millions need faster approximate methods.

AI Glossary / AI and machine learning / Advanced

Agglomerative clustering

Definition

Agglomerative clustering is a method that groups data by merging: it starts with every item as its own tiny cluster, then repeatedly joins the two closest clusters until everything is connected in one family tree of groupings.

What it means

Most clustering methods ask you to decide the number of groups in advance. Agglomerative clustering builds the whole hierarchy instead: from every customer alone, through natural small groups, up to one big group, and lets you cut the tree at whatever level is useful.

The method is bottom-up. At each step it finds the two clusters that are closest, by a chosen measure of closeness, and merges them. Repeat hundreds of times and you have a tree, called a dendrogram, showing exactly which items grouped first and which joined only at the end.

The closeness rule shapes the result. Some rules care about the nearest pair of members, producing long chain-like clusters; others care about the farthest pair, producing tight compact ones. Choosing the rule is choosing what kind of group you want.

The tree is the product. Cutting it high gives a few broad segments; cutting it low gives many fine ones. Analysts read the tree to see the data's natural structure rather than imposing a number on it.

For a business, agglomerative clustering is the exploratory tool: when you suspect your customers, products or documents have natural groupings but do not know how many, it shows you the structure before you name the segments.

How it works

The algorithm needs one input from you: a distance measure between items, such as how different two customers' purchase histories are. Everything else follows mechanically.

Each merge chooses the pair of clusters with the smallest linkage distance. Single linkage uses the closest pair of members across the two clusters; complete linkage uses the farthest pair; average linkage averages all cross-pairs. The linkage choice is the main tuning decision.

Merges are recorded in the dendrogram, where the height of each join shows how far apart the merged clusters were. Big vertical gaps suggest natural cut points: the places where merging would force unlike groups together.

The cost is computation: comparing everything to everything grows quickly, so the method suits thousands of items rather than millions, where faster approximate methods take over.

Three practical examples

Customers: group accounts by purchasing pattern and let the tree reveal whether the base is three broad types or a dozen niches.
Documents: cluster support tickets so recurring themes emerge as branches, then name the themes for routing.
Products: group SKUs by sales-behavior similarity to find which items genuinely move together for planning.

Think of it this way

Think of seating guests at a wedding by building a family tree in reverse. First the closest friends share tables, then the tables combine into sections, then the sections into sides. The seating chart is a cut through the tree at the table level.

The limit: the tree records similarity, not meaning. Two guests may sit together because they both know the groom, which is a grouping fact, not a description of who they are. Naming segments still takes judgment.

Worked example

discovering customer segments

Suppose a wholesaler believes it has four customer types but has never tested the belief. The team computes distances between customers from order history and runs agglomerative clustering.

The dendrogram shows two big early branches and several small late ones, not four clean groups. Cutting at the natural gap yields six segments, two large and four niche, which the sales team recognizes immediately and renames with real labels. The scenario is invented; reading the tree's gaps is the technique.

The four-type assumption was not wrong so much as flattened: two of the assumed types were one, and two niches had been invisible.

Segment counts in the planning deck now match structure in the data, and the tree stays in the appendix as the evidence.

Illustrative case study

A fictional software company clusters its feature requests to plan the year. The dendrogram produces one huge branch and many tiny ones, which the team reads as one big theme plus noise.

A product manager notices the huge branch mixes two audiences, enterprise and self-serve, whose requests use the same words for different needs. Splitting the distance measure by audience produces two clean trees and a defensible roadmap.

The invented lesson: clustering reflects the distances you feed it. When a cluster looks meaningless, the distance measure is usually what needs rethinking.

Common misconceptions and mistakes

It finds the true segments. It finds structure in the chosen distances; different distance measures produce different truths.
The biggest cluster is the most important. Big branches often just mean vague similarity; the small tight branches may hold the valuable niches.
You must know the segment count first. That is exactly what this method does not require; the tree shows the options.

Business use

Use it for exploration on moderate data: customer bases, product catalogs, document collections, any place where natural groupings are suspected but unknown.

Pair it with domain review: the tree proposes, the team names and validates. Then check whether the segments differ on outcomes you care about before reorganizing anything around them.

Check your understanding

A clustering produces one huge branch that mixes two audiences. Is the algorithm broken, and what should change?

Answer: No; the distance measure ignored the audience dimension. Redefine distances (or split by audience) so the structure reflects what makes requests actually differ.

Related terms

Learn next

AI for Business Owners covers using data structure to make marketing and planning choices: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Data structure

Sources

[S1] Russell and Norvig, AIMA glossary. [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Agentive logic Next: Aggregate (data warehouse)