La vitrine de diffusion des publications et contributions des chercheurs de l'ÉTS

Adaptive selection of ensembles for imbalanced class distributions


Téléchargements par mois depuis la dernière année

Plus de statistiques...

Radtke, Paulo V. W. et Granger, Éric et Sabourin, Robert et Gorodnichy, Dmitry. 2012. « Adaptive selection of ensembles for imbalanced class distributions ». In 21st International Conference on Pattern Recognition (ICPR) (Tsukuba, Japan, Nov. 11-15, 2012), p. 2980-2984. Piscataway, NJ : Institute of Electrical and Electronics Engineers Inc..

Granger E. 2012 5090 Adaptive selection of ensembles for imbalanced class distributions.pdf

Télécharger (496kB) | Prévisualisation


Boolean combination (BC) techniques have been shown to efficiently integrate the responses of multiple classifiers in the ROC space for improved accuracy and reliability. Although the impact on classification performance of imbalanced class distributions may be addressed using ensemble-based techniques, it is difficult to observe with ROC curves. Given a false alarm rate and class imbalance, performing BC in the Precision-Recall Operating Characteristic (PROC) space can lead to a higher level of performance. In practice, class distributions often change over time, and BCs should adapt to reflect operational conditions. Thus, this paper proposes an adaptive system that initially uses skewed data to generate several BCs in the PROC space. Then, during operations, the class imbalance is periodically estimated, and used to estimate the most accurate BC of classifiers among operational points of these curves. Simulation results indicate that this approach maintains a level of accuracy that is comparable to full Boolean re-combination, but for a significantly lower computational cost.

Type de document: Compte rendu de conférence
Granger, Éric
Sabourin, Robert
Affiliation: Génie de la production automatisée
Date de dépôt: 24 juill. 2013 20:43
Dernière modification: 28 janv. 2016 22:55
URI: http://espace2.etsmtl.ca/id/eprint/5090

Actions (Authentification requise)

Dernière vérification avant le dépôt Dernière vérification avant le dépôt