It has been demonstrated that kernel machines, like Multilayer Perceptron (MLP) networks, are able to classify any set of patterns defined in real domains. They can also determine the relationship occurring among the input signals, provided that a very large number of hidden neurons is included. Unfortunately this leads to a tradeoff in terms of computational resources, since the bigger is the size of the net the longer is the time needed to re-train it. This poses a problem in Magnetic Confinement Fusion devices like JET, where different operating scenarios have to be explored, requiring periodic retraining of the nets. Moreover, too complex MLP networks can be difficult to interpret and present lower generalization potential. This paper presents a new approach based on hybrid Geometrical Kernel Machines, used as regressors, to provide a functional relationship among the input variables. The particular application described is the transition between the L and H modes of confinement, with the aim of deriving a data-driven functional expression for the L-H threshold to be used for both prediction and interpretation