Romain Couillet est professeur au laboratoire LSS de CentraleSupélec à Gif. Il est invité à parler au séminaire du laboratoire MAS coencadré par la Fédération de Mathématiques de CentraleSupélec.
Titre : Large dimensional spectral clustering for data and graphs.
Résumé :
In this talk, we shall consider the generic question of spectral clustering methods for large dimensional datasets and graphs, and notably Laplacian-based community detection on heterogeneous dense graphs and kernel-based spectral clustering of data. Using advanced methods from random matrix theory, we will show that peculiar behaviors arise in these examples that disrupt with the conventional finite dimensional intuitions. New methods are then proposed to improve the performances of these traditional techniques.