Archetypal Analysis

[aleixalcacer]

Describe the workflow you want to enable

Archetypal analysis is similar to clustering analysis. However, instead of cluster centers, it seeks extremal points in the dataset (called "archetypes"), sometimes providing more interpretable results than clustering.

Describe your proposed solution

Implement the algorithm proposed in Cutler and Breiman, 1994. Since it is similar to clustering analysis algorithms, it will fit well within the fit, predict/transform interface.

In the R ecosystem there is already a specific package for archetypal analysis: archetypes

What do you think about adding this new algorithm? I can open a PR with a proposal :)

[jnothman]

I think this is much more comparable to decomposition than to clustering. The seminal paper has a relatively modest number of citations given its age, as do the papers cited in CRAN's archetypes, so I am not immediately convinced that this will be widely used by the Scikit-learn community. Arguing for it will require an example robustly demonstrating its usefulness for some machine learning problem. Even then, it may be a better candidate for scikit-learn-extra or another external package within scikit-learn's orbit.

[aleixalcacer]

Yes, it lies somewhere between decomposition and clustering. For example, if the observations are grouped by the closest archetype, archetypal analysis can be used in clustering problems. Here you can see an example of the usefulness of archetypal analysis.

Anyway, If you are still not 100% convinced, I think the best option is to create an external package. Therefore, other algorithms that do not satisfy the scikit-learn requirements can also be implemented. There are some guidelines or something like that to develop a package following the scikit-learn rules? I have only found this template, but I don't know if it is the right way to start.