Does bisecting imply equal sizes?

No; each split creates two clusters with potentially different sizes.

No; it splits groups rather than merging them.

Does it determine the ideal cluster count automatically?

The usual procedure takes a requested target count.

AI Glossary / Statistics and probability / Advanced

Bisecting K-Means

Definition

Bisecting K-Means is a divisive clustering method that repeatedly splits one existing cluster into two using K-Means until the requested number of clusters is reached. It begins with all observations together and creates a hierarchy of progressively smaller groups. The final clusters are the leaves of the resulting split tree.

What it means

Ordinary K-Means optimizes a partition with all requested centroids present. Bisecting K-Means builds the partition through successive local two-cluster problems. The choice of which existing cluster to split therefore matters, alongside the quality of each two-way split.
Bisecting means dividing into two groups, not necessarily two equally sized groups. The features and their scaling determine what similarity means.

How it works

Start with one cluster containing the observations. Choose a current cluster according to a splitting strategy, fit a two-centroid K-Means partition within it, and replace it with the two children. Each successful split increases the number of current clusters by one. Continue until the target count is reached.
Two common selection strategies are the largest cluster by observation count and the cluster with the largest within-cluster sum of squared errors, also called inertia. These strategies need not select the same cluster.
The two-way fit can use multiple initializations, retaining the best candidate split under its objective. The hierarchy constrains later splits to their parent groups; it does not perform unrestricted global reassignment among all final clusters. Initialization, stopping settings and feature preparation still influence the result.

Three practical examples

Document exploration: an analyst repeatedly divides document vectors into smaller groups, then checks representative documents before assigning descriptive names.
Operational segmentation: a team groups comparable numeric profiles into a broad-to-fine hierarchy and examines whether the divisions correspond to useful differences.
Algorithm comparison: researchers compare bisecting and ordinary K-Means on the same prepared features, reporting both computational effort and the resulting partitions.

Think of it this way

Imagine organizing a collection by repeatedly choosing one box and sorting its contents into two new boxes. You can choose the fullest box or the box whose contents are most spread out. Once items have gone into different branches, later sorting happens inside those branches.

Worked example

Consider an original one-dimensional dataset with six equally weighted observations: 0, 0, 1, 1, 10 and 14. The repeated values represent separate observations. Suppose the first two-way fit gives cluster A = {0, 0, 1, 1} and cluster B = {10, 14}. Their means are 0.5 and 12.
A has four observations and inertia 4 times 0.5 squared = 1. B has two observations and inertia (10-12) squared + (14-12) squared = 8. The combined inertia is 9. The midpoint between the means is 6.25, so assigning each observation to its nearest one of these two means preserves the stipulated split.
Suppose the target is three clusters. Under the largest-cluster strategy, A is selected because four observations exceed two. Splitting A into {0, 0} and {1, 1} gives means zero and one, each with zero inertia. B remains unchanged with inertia 8. The final three-cluster inertia is therefore 8, and the cluster sizes are two, two and two.
Under the biggest-inertia strategy, B is selected instead because its inertia 8 exceeds A's inertia 1. Splitting B into {10} and {14} gives two singletons, each with zero inertia. A remains unchanged, so the final inertia is 1. The cluster sizes are now four, one and one.
The second strategy achieves lower inertia in this example, but the first produces equal group sizes. Neither fact alone identifies the partition a particular application needs. The example does not establish that one strategy always dominates.
For comparison, the original single cluster has mean 26/6 = 13/3. Its inertia can be calculated as the sum of squared values minus six times the squared mean: 298 - 6(13/3)^2 = 556/3, approximately 185.3333. The first split reduced that value to 9. The second split then reduced it to either 8 or 1, depending on the strategy. These are objective values for the specified partitions, not guarantees about every initialization.
Now multiply every input value by ten. The first two means become 5 and 120. Squared distances increase by a factor of one hundred, so A's and B's inertias become 100 and 800. Their ordering is unchanged, as are their observation counts. Scaling every coordinate by the same positive constant preserves these comparisons; scaling only one feature in a multidimensional problem can change the geometry and the selected splits.
Starting from one cluster, two successful splits were needed to obtain three clusters. Continuing to split cannot undo the first separation between A and B. All calculations here use explicit candidate partitions rather than measured software performance.

Illustrative case study

Fictional company Maple Analytics explores numeric profiles from a simulated workflow. Its team prepares the features consistently, records the target cluster count and compares the two selection strategies. Analysts inspect both cluster sizes and representative observations.
When a low-inertia result creates small groups, they check whether those groups are interpretable and stable across reasonable settings. They avoid turning cluster labels directly into business rules without further evaluation.

Common misconceptions and mistakes

A bisection need not create equal halves. Choosing the cluster with the most observations is different from choosing the one with the most inertia, as the example demonstrates.
More clusters generally allow lower squared error, so comparing inertia across different cluster counts alone does not settle model selection. A further mistake is assuming that local two-way solutions guarantee the globally best final partition. Outliers, unsuitable feature scales and shapes poorly represented by centroid-based groups can still cause misleading results.

Business use

Bisecting K-Means offers a way to explore hierarchical segmentations while working on subsets during successive splits. Evaluation should connect the resulting groups to the intended use and compare stability, interpretability and computation. A numerical reduction in inertia is useful evidence about compactness, not direct evidence of commercial value.

Check your understanding

Why did the strategies choose different groups?

Answer: A had more observations, while B had greater inertia.

What was the lower final inertia?

Answer: One, after splitting B.

Why can an early split matter later?

Answer: Subsequent divisions remain within its branches.

Related terms

k-means
Divisive clustering
Hierarchical clustering
Centroid
Inertia
Feature scaling
Cluster validation

Learn next

Repeat the six-point example with a different pair of distant observations and identify when cluster size and inertia recommend different next splits.

Sources

Scikit-learn, clustering user guide, Bisecting K-Means: https://scikit-learn.org/stable/modules/clustering.html#bisecting-k-means . Scikit-learn, BisectingKMeans API: https://scikit-learn.org/stable/modules/generated/sklearn.cluster.BisectingKMeans.html . The six-point calculations and company scenario are original illustrations.

Previous: Birnbaum's theorem Next: Biserial correlation