Handling Imbalanced Regression Tasks through Utility-based Regression

Colloque / Congrès / Forum Public-cible: Ouvert au grand public
Several important application domains require predicting a numeric variable with a very imbalanced distribution. Frequently, the end-user is particularly interested in the performance of the models over a range of values of this target variable that is under-represented in the available data set (e.g. forecasting some unusual readings of some industrial sensor). This creates serious problems not only on how to properly evaluate the performance of alternative models, but also on how to obtain the "optimal" models from this type of data. In this seminar we will present the concept of utility-based regression that was proposed to handle tasks where the preference bias of the end-user is not uniform across the domain of the target variable. Imbalanced regression can be seen as an instance of these problems. We will describe the concepts associated with utility-based regression and also some concrete proposals for addressing imbalanced regression based on these concepts.
20.06.2017 14:00 - 16:00
Site PER 21 / Salle C120
Bd de Pérolles 90, 1700 Fribourg
Département d'Informatique
Prof. Philippe Cudré-Mauroux
Bd de Pérolles 90
1700 Fribourg
026 300 83 22
Prof. Dr. Luis Torgo, University of Porto / INESC Tec LA
Pièces jointes