Estimation of the conditional risk in classification: the swapping method

Jean-Jacques Daudin (1) , Tristan Mary-Huard (1)

1 MIA-Paris - Mathématiques et Informatique Appliquées

The bias of the empirical error rate in supervised classification is studied. It is shown that this bias can be understood as a covariance between the classification rule and the labeling of the training data. From this result, a new penalized criterion is proposed to perform model selection in classification. Applications of the resulting algorithm to simulated and real data are presented.

Mots clés: classification; model selection; covariance penalty

Domaines: Statistiques [math.ST]

Jean-Jacques Daudin, Tristan Mary-Huard. Estimation of the conditional risk in classification: the swapping method. Computational Statistics and Data Analysis, 2008, 52 (6), pp.3220-3232. ⟨10.1016/j.csda.2007.10.003⟩. ⟨hal-01197609⟩