Methods for Label Ranking
Intervenant(s) : Eyke Hüllermeier (University of Marburg)
This talk is devoted to a relatively novel machine learning learning problem called label ranking. This problem can be seen as a generalization of conventional classification learning. Instead of learning a classifier that maps instances to single class labels, the goal is to learn a "label ranker" in the form of a mapping from instances to rankings over the complete set of class labels. The talk will give an overview of methods for label ranking that have been proposed in the recent literature and discuss two of these methods in more detail. One of them, called ranking by pairwise comparison, first induces a binary preference relation from suitable training data using a natural extension of pairwise classification. A ranking is then derived from the preference relation thus obtained by means of a ranking procedure. The second method is an instance-based approach built upon the nearest neighbor estimation principle. It makes use of a probability model on rankings, which is known as the Mallows model in statistics, and derives predictions in the form of maximum likelihood estimates.
Thomas.Baerecke (at) nulllip6.fr