It is related to exploratory grouping, but observations can be represented by mixtures rather than exclusive assignments.

Must archetypes be real records?

No. The classical method permits convex mixtures of records.

Does the method discover the correct number automatically?

The number is a modeling choice that needs evaluation and interpretation.

AI Glossary / Data science and analytics / Advanced

Archetypal analysis

Definition

Archetypal analysis is an unsupervised method that represents observations as mixtures of a small number of characteristic profiles called archetypes. Each archetype is itself a mixture of observed data points. In the classical formulation, both sets of mixture weights are nonnegative and sum to one, and the method seeks to minimize reconstruction error.

What it means

Instead of assigning every observation to one typical group, the method describes how an observation blends several contrasting profiles. A learning session might combine features of an example-focused profile and a reference-focused profile, rather than belonging entirely to one category.
The word archetype does not mean a permanent human personality or an ideal customer. It is a mathematical profile derived from selected measurements. Archetypes depend on the dataset, feature preparation and chosen number of profiles. Naming them can aid interpretation, but a memorable label should not be allowed to replace the underlying measurements or imply traits that were never recorded.

How it works

Choose the variables and the number of archetypes. The classical method alternates between estimating how observations mix the archetypes and how archetypes mix the observations, reducing the sum of squared reconstruction errors under the weight constraints.
Because the weights are nonnegative and sum to one, these are convex combinations. The archetypes stay within the convex hull of the observed data, while approximations lie within the hull formed by the selected archetypes. This differs from freely combining principal-component directions with positive or negative coefficients.
Feature scaling influences squared distances, so choose it deliberately. Try multiple starting points and examine stability rather than assuming one fitted solution is uniquely correct. Assess the trade-off between a smaller, interpretable representation and a closer reconstruction of the data.

Three practical examples

A learning publisher explores session profiles using proportions of time spent on definitions, examples and exercises. Mixed weights help describe sessions that cross these activities without forcing them into exclusive labels.
An operations team summarizes daily workload profiles across several service channels. It examines combinations of contrasting patterns when planning scenarios, while checking whether rare operational incidents distort the profiles.
A product team studies usage profiles across selected features. It uses archetypes to organize exploratory discussions, then checks proposed design changes with users instead of treating the profiles as proven user needs.

Think of it this way

Imagine describing paint colors as mixtures from a small palette, with every mixture using nonnegative amounts that add to the whole. Archetypal analysis also constructs its palette from mixtures of colors actually observed. The analogy captures the two mixture constraints. Real behavioral data, however, may not be reconstructed exactly by a small palette, and the chosen variables determine what the colors represent.

Worked example

Consider three observed profiles with two measured features: P = (10, 0), Q = (0, 10) and R = (6, 4). Choose two archetypes Z1 = P and Z2 = Q. Each archetype is a valid convex combination of observations because it places weight 1 on one observation and 0 on the others.
R is reconstructed as 0.6 × Z1 + 0.4 × Z2 = (6, 4). The weights are nonnegative and sum to 1. P and Q also reconstruct exactly, so the total squared reconstruction error for these three points is zero.
Now use this fixed pair to approximate a new point S = (8, 8). Every mixture has coordinates summing to 10, so it cannot equal S, whose coordinates sum to 16. The closest mixture is (5, 5), with squared error (8 − 5)^2 + (8 − 5)^2 = 18. This invented example shows a representational limit; refitting with new data could change the archetypes.

Illustrative case study

In a fictional project, Elm Learning explores how readers use long glossary articles. Analysts describe sessions through three proportions: reading explanatory sections, studying examples and attempting exercises. They fit several small archetypal representations and examine the original sessions contributing to each profile.
An initially dramatic profile turns out to reflect a tracking error that counted an inactive browser tab as exercise time. After correcting the measurement, that profile changes substantially. The team also finds many sessions mix the remaining profiles, making rigid reader labels misleading.
Elm uses the analysis to choose a few navigation ideas for user testing. It retains the feature definitions and fitting settings with its findings. The value is an interpretable exploration of recorded activity, not a claim to have discovered fixed types of people or the reasons behind every session.

Common misconceptions and mistakes

Archetypes are not necessarily individual observations; they may be mixtures. They also differ from cluster centers that summarize dense groups. A mixture weight is a reconstruction coefficient, not automatically a probability that someone belongs to a psychological category.
Adding archetypes can change the previously fitted profiles, so do not assume a three-profile solution simply appends one profile to a two-profile solution. Outliers, feature units and measurement errors can influence the result.
Low reconstruction error alone does not establish business usefulness. A highly detailed representation can be difficult to interpret or unstable across datasets, and labels can exaggerate distinctions that the data supports only weakly.

Business use

Use archetypal analysis for exploratory summaries when contrasting profiles and mixtures are useful to decision-makers. Document the observations, features and preparation choices before presenting named archetypes.
Review reconstruction quality, stability and interpretability together. Keep an appropriate comparison with simpler summaries, and examine whether results persist across relevant time periods. If profiles inform a product or service change, evaluate that change separately. For people-related data, communicate that the profiles describe measured activity in context rather than an individual's inherent identity or future behavior.

Check your understanding

What makes the mixture convex?

Answer: Nonnegative weights that sum to one.

Why might scaling change the result?

Answer: It changes the relative contribution of variables to reconstruction error.

What does a 0.6 weight mean in the example?

Answer: That archetype contributes 60% of the fitted convex mixture, not a 60% personality probability.

Related terms

Learn next

Draw a small set of two-dimensional observations and compare a two-archetype reconstruction with cluster centers. Then change feature scaling and inspect how the fitted representation changes.

Sources

Cutler and Breiman, Archetypal Analysis, Technometrics, 1994: original formulation and alternating optimisation. https://stat.cmu.edu/technometrics/90-00/vol-36-04/v3604338.pdf Coresets for Archetypal Analysis, NeurIPS 2019: later primary research describing the constrained representation. https://papers.nips.cc/paper_files/paper/2019/file/7f278ad602c7f47aa76d1bfc90f20263-Paper.pdf The numerical construction and Elm Learning case are original teaching illustrations, not reported research findings.

Previous: Arc standard Next: Architecture