Case-based Multilabel Ranking

Klaus Brinker, Eyke Hüllermeier

We present a case-based approach to multilabel ranking, a recent extension of the well-known problem of multilabel classification. Roughly speaking, a multilabel ranking refines a multilabel classification in the sense that, while the latter only splits a predefined label set into relevant and irrelevant labels, the former furthermore puts the labels within both parts of this bipartition in a total order. We introduce a conceptually novel framework, essentially viewing multilabel ranking as a special case of aggregating rankings which are supplemented with an additional virtual label and in which ties are permitted. Even though this framework is amenable to a variety of aggregation procedures, we focus on a particular technique which is computationally efficient and prove that it computes optimal aggregations with respect to the (generalized) Spearman rank correlation as an underlying loss (utility) function. Moreover, we propose an elegant generalization of this loss function and empirically show that it increases accuracy for the subtask of multilabel classification.