Clusterwise methods, past and present

Gilbert Saporta (1)

1 CEDRIC - MSDMA - CEDRIC. Méthodes statistiques de data-mining et apprentissage

Instead of fitting a single and global model (regression, PCA, etc.) to a set of observations, clusterwise methods look simultaneously for a partition into k clusters and k local models optimizing some criterion. There are two main approaches: 1. the least squares approach introduced by E.Diday in the 70's, derived from k-means 2. mixture models using maximum likelihood but only the first one easily enables prediction. After a survey of classical methods, we will present recent extensions to functional, symbolic and multiblock data.

Mots clés: clusterwise regression; mixture models; dimension reduction; PLS regression

Domaines: Statistiques [stat]

Gilbert Saporta. Clusterwise methods, past and present. ISI 2017 61st World Statistics Congress, Jul 2017, Marrakech, Morocco. ⟨hal-02473529⟩