Abstract. We propose a novel active learning strategy based on the compression framework of [9] for label ranking functions which, given an input instance, predict a total order over a predefined set of alternatives. Our approach is theoretically motivated by an extension to ranking and active learning of K