Case-Based Label Ranking

Abstract

Label ranking studies the problem of learning a mapping from instances to rankings over a predefined set of labels. We approach this setting from a case-based perspective and propose a sophisticated k -NN framework as an alternative to previous binary decomposition techniques. It exhibits the appealing property of transparency and is based on an aggregation model which allows one to incorporate a variety of pairwise loss functions on label rankings. In addition to these conceptual advantages, we empirically show that our case-based approach is competitive to state-of-the-art model-based learners with respect to accuracy while being computationally much more efficient. Moreover, our approach suggests a natural way to associate confidence scores with predictions, a property not being shared by previous methods.