Prof. Johannes Fürnkranz
Knowledge Engineering Group
TU Darmstadt
http://www.ke.tu-darmstadt.de/staff/juffi

Preference Learning by Pairwise Decompositions

Preference Learning is a recent learning setting, which may be viewed
as a generalization of several conventional problem settings, such as
classification, multi-label classification, ordinal classification, or
label ranking. In this talk, I will give a brief introduction into
this area and then focus on the learning by pairwise comparisons
approach. From a machine learning point of view, this approach is
especially appealing as it decomposes complex prediction problems into
a number of simpler learning problems, each one comparing a particular
pair of options. Its main advantages lie in the simplicity of the
decision boundaries of the resulting binary learning problems, as well
as in the flexibility in the choice of an aggregation function which
allows to minimize different loss functions without the need for
re-training. At first sight, a key disadvantage seems to be the large
number of resulting classifiers (one for each pair of labels), but it
can be shown that training and testing time can be reduced to (almost)
linear in the number of labels, so that large problems may be tackled
efficiently.