Nicolas Vayatis
http://nvayatis.perso.math.cnrs.fr/
Wednesday 13th March 2013
Time: 4pm
Basement Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
New perspectives on the problem of scoring high dimensional data
We consider the scoring approach for the inference of an order relationship from high dimensional data when ordinal labels (ratings) are available. In this setting, few algorithms are known with theoretical guarantees for consistency and well-posedness. In the talk, key issues will be addressed such as the functional character of the performance measure (e.g. ROC surfaces) and the principle of aggregation of decision rules when local averaging is meaningless. We will see how recent advances on optimality and aggregation for nonparametric scoring opens new perspectives for the design of efficient algorithms.