Label noise detection under the noise at random model with ensemble filters

Label noise detection has been widely studied in Machine Learning because of its importance improving training data quality. Satisfactory achieved by adopting ensembles classifiers. In this approach, an instance is assigned as mislabeled if a high proportion members the pool misclassifies it. Previous authors have empirically evaluated approach; nevertheless, they mostly assumed that label generated completely at random dataset. This strong assumption since other types are feasible practice and can influence results. work investigates performance ensemble under two different models: Noisy Random (NAR), which probability depends on class, comparison to Completely model, entirely independent. setting, we investigate effect class distribution it changes total level observed dataset NAR assumption. Further, evaluation vote threshold conducted contrast with most common approaches literature. many performed experiments, choosing generation model over another lead results when considering aspects such imbalance ratio among classes.