Uncertainty-Aware Resampling Method for Imbalanced Classification Using Evidence Theory
Résumé
Class imbalance is a common issue in many real world classification problems. It refers to situations where the number of observations in the training dataset significantly differs for each class. Ignoring this issue will make it more challenging for classifiers to properly learn data characteristics, which results in poor performance. Many strategies have been proposed to deal with this issue. The most common one is tackling the imbalance at the preprocessing level, by re-sampling the training set. However, imbalanced classification can be affected by other data factors, such as uncertainty, i.e., ambiguous samples and noise. In this paper, we propose an uncertainty-aware hybrid resampling technique based on the theory of evidence to tackle imbalanced binary datasets in the presence of aleatoric uncertainty. A soft evidential structure is assigned to each object in the training set, which is later used to clean the dataset out of overlapping and noisy majority samples, and then selectively generate synthetic minority objects using a modified SMOTE algorithm. Experimental results on benchmark imbalanced datasets have shown significant improvement over popular re-sampling techniques.
Origine : Fichiers produits par l'(les) auteur(s)