On predictive accuracy and risk minimization in pairwise label ranking

We study the problem of label ranking, a machine learning task that consists of inducing a mapping from instances to rankings over a finite number of labels. Our learning method, referred to as ranking by pairwise comparison (RPC), first induces pairwise order relations (preferences) from suitable training data, using a natural extension of so-called pairwise classification. A ranking is then derived from a set of such relations by means of a ranking procedure. In this paper, we first elaborate on a key advantage of such a decomposition, namely the fact that it allows the learner to adapt to different loss functions without re-training, by using different ranking procedures on the same predicted order relations. In this regard, we distinguish between two types of errors, called, respectively, ranking error and position error. Focusing on the position error, which has received less attention so far, we then propose a ranking procedure called ranking through iterated choice as well as an efficient pairwise implementation thereof. Apart from a theoretical justification of this procedure, we offer empirical evidence in favor of its superior performance as a risk minimizer for the position error.