Abstract (EN):
Ensemble methods have been receiving an increasing amount of attention, especially because of their successful application to problems with high visibility (e.g., the NetFlix prize, Kaggle competitions, etc). An important challenge in ensemble learning (EL) is the management of a set of models in order to ensure accuracy and computational efficiency, particularly with a large number of models in highly dynamic environments. We plan to use metalearning (MtL) to improve the performance of one of the most important EL algorithms: bagging. MtL uses data from past experiments to build models that relate the characteristics of learning problems with the behaviour of algorithms. Our approach consists in using MtL techniques that act at the level of 1) ensemble pruning and 2) ensemble integration to improve the performance of the original bagging algorithm. On the one hand, we present results of a technique that allows to prune bagging ensembles before actually generating the individual models with a performance equal to the original algorithm. On the other hand, we expose how we plan to extend the work done in 1) to include a dynamic approach. Our final goal is to achieve a MtL method that is able to select the most suitable subset of models according to the characteristics of the instance that is being predicted.
Idioma:
Inglês
Tipo (Avaliação Docente):
Científica