Is the first split at the median?

No. The method uses the arithmetic mean, so the groups need not contain equal numbers of observations.

Does it prove the data follow a power law?

No. Distribution fitting and testing are separate tasks.

Is the number of classes fixed in advance?

Usually it emerges from the recursive procedure and its stopping rule rather than being supplied as a fixed class count.

AI Glossary / Data engineering and architecture / Intermediate

Head/tail breaks

Definition

Head/tail breaks is a classification method that repeatedly divides data around their arithmetic mean, focusing the next division on the smaller group of larger values. Introduced by Bin Jiang, it is intended to reveal hierarchy in strongly uneven distributions with many small values and relatively few large ones. It is a classification scheme, not a test that proves a particular probability distribution.

What it means

A dataset of settlement sizes or street lengths may contain many modest values and a few very large ones. Equal-width or equal-count classes can emphasize different aspects of that pattern. Head/tail breaks follows the large-value group recursively, producing boundaries related to the data's repeated imbalance. In this terminology, the head contains the larger values, while the tail contains the more numerous smaller values. That naming should not be confused with every statistical use of the word tail.

How it works

Compute the mean and separate values above it from the remaining values using a stated boundary convention. If the large-value head satisfies the chosen continuation rule, repeat on that head. Stop when the required imbalance no longer holds or further division is not meaningful. Implementations can specify thresholds and stopping details, including treatment of equal values and very small groups. Report those choices because they affect the resulting number of classes and boundaries.

Three practical examples

A cartographer classifies highly uneven settlement sizes to make the larger-size hierarchy visible.
A network analyst explores a skewed distribution of node-related measures, while checking what the resulting classes actually represent.
An AI data analyst examines uneven document lengths or usage counts as an exploratory display, without treating the classes as validated targets for prediction.

Think of it this way

Think of separating a collection of stones at the average weight, then examining the smaller pile of heavier stones in the same way. The process focuses on successive levels among the large values. The analogy is limited because a numerical classification does not explain how the stones formed or prove that their weights follow a specific statistical law.

Worked example

Take the illustrative values 1, 1, 1, 2, 2, 3, 4, 8, 16 and 32. Their sum is 70, so the first mean is 7. The head is 8, 16 and 32: three of ten values. Its mean is 56/3, about 18.67, leaving 32 as the next head. Under a toy continuation rule allowing a head of at most 40% and stopping at a singleton, these divisions yield boundaries at 7 and about 18.67. The arithmetic demonstrates the recursion; ten selected numbers do not establish a heavy-tailed population distribution.

Illustrative case study

Fictional case: Rowan Maps compares several classification schemes for a skewed count dataset. Head/tail breaks makes distinctions among the largest values clearer, but the team also displays the underlying distribution and explains the chosen stopping rule. When comparing a later year, it considers whether changing class boundaries would obscure change over time. This fictional case illustrates display choices rather than a claim that one classification is always the most truthful.

Common misconceptions and mistakes

Do not treat the method as proof of a power law, fractality or causal hierarchy. Different uneven distributions can produce similar-looking classes. Do not assume a 40% threshold is a universal mathematical definition of heavy tails; it is an implementation convention used in some versions. Outliers, missing values and measurement errors can affect means and boundaries. Also, classes calculated separately for two datasets may not be directly comparable even if they use the same colors or class labels.

Business use

Businesses may use head/tail breaks for exploratory visualization when a small number of large values dominate a measure. Start by checking data quality and the distribution rather than selecting the method for visual drama. Label class boundaries clearly and preserve access to the underlying values. For AI, avoid turning exploratory bins into ground-truth categories without a task-specific reason. If the classes inform decisions, assess sensitivity to the stopping rule and to small changes in the data.

Check your understanding

A classification produces several levels among large values. Can a report conclude that the underlying process is fractal solely from that result?

Answer: No. The classification alone does not establish such a structural or statistical claim.

Related terms

data classification
arithmetic mean
skewed distribution
heavy tail
choropleth map
quantile classification
natural breaks

Learn next

Apply the first two mean splits to a small uneven dataset and record the head proportion at each step. Compare with equal-count classes. Then change the largest value and observe how sensitive the boundaries are before choosing a display for a real audience.

Sources

Primary author reference: Bin Jiang, Head/tail Breaks: A New Classification Scheme for Data with a Heavy-tailed Distribution, https://arxiv.org/abs/1209.2801 and https://arxiv.org/pdf/1209.2801 . Original-paper abstract and method discussion consulted. Numerical example states its own stopping convention; broader distribution claims are not inferred. Full implementation and editorial review remain pending.

Previous: Head Next: Healing brush