In this article we show how the structure of an autoencoder network can be used to produce the output of an archetypal analysis reducing the computational effort.
Archetypal Analysis as an autoencoder
Archetypal Analysis (AA) is a data approximation method which approximates data points by convex (positive coefficients with unitary sum) combination of prototypes, denoted as 'archetypes'. 'Archetypes' themselves are a convex combination of the original data points.
AA was introduced as a dimensional reduction method in [1, 2] with more interpretable results when compared to Principal Component Analysis (PCA). The original comparison study had the goal of designing facemasks for the Swiss Army, starting from a dataset consisting of 6 head dimensions for 200 soldiers.
Please complete the following form to access to the free download
“Archetypal analysis as an autoencoder”