We present an algorithm which uses example pairs of equal or unequal class labels to select a projection on a kernel-induced Hilbert space. A representation of .nite dimensional projections as bounded lin- ear functionals on a space of Hilbert-Schmidt operators is exploited to give bounds on the Rademacher complexity of the class of hypotheses searched, leading to PAC-type performance guarantees for the resulting feature maps. The proposed algorithm returns the projection onto the span of the principal eigenvectors of an empirical operator constructed in terms of the example pairs. Experiments demonstrate an e¤ective trans- fer of knowledge between di¤erent but related learning tasks.