Machine learning tools have been used since a long time to study disruptions and to predict their occurrence. On the other hand, the challenges posed by the quality and quantities of the data available remain substantial. In this paper, methods to optimize the training dataset and the potential of advanced machine learning tools, based on kernels, are explored and assessed. Various alternatives, ranging from appropriate selection of the weights to the inclusion of artificial points, have been investigated in order to improve the quality of the training dataset. Support Vector Machines (SVM), Relevance Vector Machines and one class SVM have been compared. The relative performances of the different approaches are first assessed using synthetic data. Then they are applied to a relatively large database of JET disruptions. It is shown that in terms of final results, the optimization of the training databases proved to be very productive. On the other hand, for the problem of disruption prediction, the two classes SVM remains the most performing machine learning tool that were tested in this contribution.