Criteria for variable selection with dependence

Aurélie Boisbunon (1) , Stephane Canu (1) , Dominique Fourdrinier (1)

1 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes

We propose a new data-driven procedure of model selection based loss estimation and valid for the whole family of spherically symmetric distributions, allowing for dependence between noise components. We give explicit formulas for the Firm Shrinkage estimator, and we compare by simulation our results in terms of selection with the classical AIC, BIC and leave-one-out cross-validation.

Domains: Statistics [math.ST]; Statistics Theory [stat.TH]; Machine Learning [cs.LG]

Aurélie Boisbunon, Stephane Canu, Dominique Fourdrinier. Criteria for variable selection with dependence. 2011. ⟨hal-00626307⟩