High-dimensional variables clustering based on sub-asymptotic maxima of a weakly dependent random process
Résumé
The dependence structure between extreme observations can be complex. For that purpose, we see clustering as a tool for learning the complex
extremal dependence structure. We introduce the Asymptotic Independent block (AI-block) model, a model-based clustering where population-
level clusters are clearly defined using independence of clusters’ maxima of a multivariate random process. This class of models is identifiable
allowing statistical inference. With a dedicated algorithm, we show that sample versions of the extremal correlation can be used to recover the
clusters of variables without specifying the number of clusters. Our algorithm has a computational complexity that is polynomial in the dimension
and it is shown to be strongly consistent in growing dimensions where observations are drawn from a stationary mixing process. This implies that
groups can be learned in a completely nonparametric inference in the study of dependent processes where block maxima are only subasymptotic,
i.e., approximately extreme value distributed.