PROJECT TITLE :

Archetypal Analysis for Nominal Observations

ABSTRACT:

Archetypal analysis is a widespread exploratory tool that explains a set of observations as compositions of few 'pure' patterns. The standard formulation of archetypal analysis addresses this problem for real valued observations by finding the approximate convex hull. Recently, a probabilistic formulation has been urged that extends this framework to other observation varieties such as binary and count. In this article we tend to any extend this framework to address the final case of nominal observations that includes, for example, multiple-choice questionnaires. We read archetypal analysis in a generative framework: this allows explicit management over choosing a suitable number of archetypes by assigning appropriate prior data, and finding economical update rules using variational Bayes'. We demonstrate the efficacy of this approach extensively on simulated knowledge, and three universe examples: Austrian guest survey dataset, German credit dataset, and SUN attribute image dataset.