Numerical performance of Penalised Comparison to Overfitting for bandwith selection in kernel density estimation

Numerical performance of Penalised Comparison to Overfitting for bandwith selection in kernel density estimation

Jeudi 29 novembre 2018, 10:15

Salle de séminaires M.0.1

Groupe de travail de Statistique
Suzanne Varet

Laboratoire de Mathématique d'Orsay, Université Paris Sud

In multivariate kernel density estimation, the bandwidth selection remains a challenge in terms of algorithmic performance and quality of the resulting estimation. A recently developped method, the Penalized Comparison to Overfitting (PCO), is compared to other usual bandwidth selection methods for multivariate and univariate kernel density estimation. In particular, the cross-validation and plug-in estimators are numerically investigated and compared to PCO. This study points out that the PCO can outperform the others classical methods without algorithmic additionnal cost.

Cet exposé rentre dans le cadre des projets ANR SMILES (ANR-18-CE40-0014) et RIN Asterics (17B01101GR).

 
Exposé