In domains like bioinformatics, information retrieval and social network\nanalysis, one can find learning tasks where the goal consists of inferring a\nranking of objects, conditioned on a particular target object. We present a\ngeneral kernel framework for learning conditional rankings from various types\nof relational data, where rankings can be conditioned on unseen data objects.\nWe propose efficient algorithms for conditional ranking by optimizing squared\nregression and ranking loss functions. We show theoretically, that learning\nwith the ranking loss is likely to generalize better than with the regression\nloss. Further, we prove that symmetry or reciprocity properties of relations\ncan be efficiently enforced in the learned models. Experiments on synthetic and\nreal-world data illustrate that the proposed methods deliver state-of-the-art\nperformance in terms of predictive power and computational efficiency.\nMoreover, we also show empirically that incorporating symmetry or reciprocity\nproperties can improve the generalization performance.\n