Gaussian Models and Kernel Methods

Overview & Synopsis

Kernel methods have been extensively used to transform initial datasets by mapping them into a so-called kernel space or RKHS, before applying some statistical procedure onto transformed data. In particular, this kind of approach has been explored in the literature to try and make some prescribed probabilistic model more accurate in the RKHS, for instance Gaussian mixtures for classification or mere Gaussians for outlier detection. Therefore this thesis studies the relevancy of such models in kernel spaces. In a first time, we focus on a family of parameterized kernels - Gaussian RBF kernels - and study theoretically the distribution of an embedded random variable in a corresponding RKHS. We managed to prove that most marginals of such a distribution converge weakly to a so-called ''scale-mixture'' of Gaussians - basically a Gaussian with a random variance - when the parameter of the kernel tends to infinity. This result is used in practice to device a new method for outlier detection. In a second time, we present a one-sample test for normality in an RKHS based on the Maximum Mean Discrepancy. In particular, our test uses a fast parametric bootstrap procedure which circumvents the need for re-estimating Gaussian parameters for each bootstrap replication.