Predicting Partial Orders: Ranking with Abstention

Abstract

The prediction of structured outputs in general and rankings in particular has attracted considerable attention in machine learning in recent years, and different types of ranking problems have already been studied. In this paper, we propose a generalization or, say, relaxation of the standard setting, allowing a model to make predictions in the form of partial instead of total orders. We interpret such kind of prediction as a ranking with partial abstention: If the model is not sufficiently certain regarding the relative order of two alternatives and, therefore, cannot reliably decide whether the former should precede the latter or the other way around, it may abstain from this decision and instead declare these alternatives as being incomparable. We propose a general approach to ranking with partial abstention as well as evaluation metrics for measuring the correctness and completeness of predictions. For two types of ranking problems, we show experimentally that this approach is able to achieve a reasonable trade-off between these two criteria.