Augmented quantization : a general approach to mixture models - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Augmented quantization : a general approach to mixture models

Résumé

Quantization methods classically provide a discrete representation of a continuous set. This type of representation is relevant when the objective is the visualisation of weighted prototype elements representative of a continuous phenomenon. Nevertheless, more complex descriptions may be investigated. In this sense, mixture models identify subpopulations in a sample, arising from different distributions. The Gaussian mixture model is particularly popular and relies on the Expectation-Maximisation (EM) algorithm [1] for maximum likelihood estimation. The computation of the likelihood limits the type of distributions in the mixture; more specifically, for the Dirac distributions and even uniform components despite their high interest in practice for processing computer experiments. Their visualization is convenient and can lead to a sensitivity analysis where variables with largest marginals are least sensitive and vice versa, as shown by our application to a flooding real case in [2]. The objective of our study is to build a very general method to provide a mixture model that approximates a sample (xi) n i=1 ∈ X n ⊂ R n from a random variable X. The representatives of the sample are the calculated components of the mixture, chosen in a parameterized family of laws denoted R. We investigate, for a given number of representatives ℓ ∈ N, the mixture X˜ ℓ = R(J) approximating X. The representatives (R(j) ) ℓ j=1 and the discrete random variable J ∈ {1, . . . , ℓ} need to be optimised.
gdr_2023_sire_charlie.pdf (1.48 Mo) Télécharger le fichier
mascotnum2023-template_SIRE.pdf (202.65 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence
Copyright (Tous droits réservés)
Origine Fichiers produits par l'(les) auteur(s)
Licence
Copyright (Tous droits réservés)

Dates et versions

hal-03975978 , version 1 (06-02-2023)

Identifiants

  • HAL Id : hal-03975978 , version 1

Citer

Charlie Sire, Rodolphe Le Riche, Didier Rullière, Jérémy Rohmer, Lucie Pheulpin, et al.. Augmented quantization : a general approach to mixture models. MASCOT-NUM2023, UQ@Paris-Saclay, Apr 2023, LE CROISIC, France. ⟨hal-03975978⟩
163 Consultations
60 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More