The classical setting of the supervised learning problem is to learn a decision rule from labeled data where data is points in some space [Xscr ] and labels are in {+1,−1}. In this paper we consider an extension of this supervised learning setting: given training vectors in space [Xscr ] along with labels and description of this data in another space [Xscr ]^*, find in space [Xscr ] a decision rule better than the one found in the classical setting [1]. Thus, in this setting we use two spaces for describing the training data but the test data is given only in the space [Xscr ]. In this paper, using SVM type algorithms, we demonstrate the potential advantage of the new setting.