Documentation / Guide méthodologique :
The sensivity measures based on Hilbert-Schmidt independence criterion (HSIC) [Gretton05,DaVeiga15] aims at analyzing the influence
of input variables
on the output variables
determining the dependance computing the dissimilary between the joint distribution and the product of the marginal distributions .
We associate respectively to and Reproducing Kernel Hilbert Space (RKHS) and associating projection functions and defined by characteristic kernels and . Data are projected in a characteristic space using the kernel trick operation.
The advantage of this approach is that the projection has not to be known, we just need to arbitrary define the kernel. Using the kernel trick operation, the HSIC measure, defined as Hilbert-Shmidt of the cross covariance operator, is expressed as
which is estimated as
with
where is the Kronecker operator Sensitivity indices are therefore defined such as
These indices are mainly used for screening or ranking the parameters.