Since the appearance of kernel methods, several researchers have been opted to combine multiple kernel instead of using singleone. While this method performance is directly related to a suite base kernels specification. The complexity of this task puts in evidence the needto make it automatic. In this context Multiple Kernel Learning (MKL) has received significant attention from researchers. In the literature, MKLis often solved by a minimization of kernels weights problem. However in this optimization problem all kernels are often used witch leads tocomplex computations. Another limitation is met in the case where the total number of kernels is high, but only a relatively small number ofthem is need to represent the target function. It is therefore a sparse problem. In this paper we wanted to improve the efficiency of MKL withless kernels, we seek so a compromise between performance, sparsity and training time complexity. We extend a recent version of MKL witchformulated as a binary classification problem with an additional constraint. This constraint ensures that the kernels be positive semi-definite. Wereformulate this optimization problem by compensating the empirical risk with a regularization term based on sparsity. Experiments on multipledata sets show the promising performance of our method compared with a competitive method.

EXPLOR_STEP=$WICRI_ROOT/Wicri/France/explor/LeHavreV1/Data/France/Analysis

HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 000379 | SxmlIndent | more

HfdSelect -h $EXPLOR_AREA/Data/France/Analysis/biblio.hfd -nk 000379 | SxmlIndent | more

{{Explor lien
   |wiki=    Wicri/France
   |area=    LeHavreV1
   |flux=    France
   |étape=   Analysis
   |type=    RBID
   |clé=     Hal:hal-01079154
   |texte=   Parcimonie et corrélation pour la sélection de noyaux
}}