Discrimination par modèles additifs parcimonieux
Avalos Marta, Yves Grandvalet and Christophe Ambroise
We propose a nonparametric classification method designed to support the interpretability of the prediction. On the one hand, the use of generalized additive models makes it possible to represent the effect of each input variable on the output variable graphically. On the other hand, parameters of this model are estimated via penalized likelihood, where the term of regularization generalizes LASSO to the splines functions. This penalization favors parsimonious solutions selecting one part of the set of input variables, while allowing a flexible modeling of the dependence on the selected variables. We study the adaptation of various analytical model selection criteria to these models, and we evaluate them on two real data sets.