Return
Antinoisy-Labeling AUC Maximization Learning for Fuzzy Classification on Imbalanced Data
DOI:10.1109/TFUZZ.2025.3642192.png)
Abstract
En 中文
By means of both strong uncertainty-handling capability of fuzzy systems and unique advantage of area under the receiver operating characteristic (ROC) curve (AUC) maximization learning for imbalanced classification, a fuzzy classifier consisting of weighted fuzzy rules can be well trained through AUC maximization learning to accomplish an efficient classification on imbalanced data. However, it will be seriously challenged in the presence of noisy labels. This study will explore a novel route for such a fuzzy classifier under AUC maximization learning framework to overcome this challenge. The basic idea is to derive a newly designed antinoisy-labeling AUC maximization learning objective with the introduction of antinoisy-labeling rate and then develop the corresponding training algorithm, which induces the corresponding fuzzy classifier as antinoisy-labeling and fuzzy classifier for imbalanced data (ANLI-FC). The distinctive novelty of the resultant fuzzy classifier ANLI-FC on imbalanced data with label noises exists in simultaneous AUC maximization learning of both antinoisy-labeling rate and all the parameters in all the weighted fuzzy rules, with theoretical guarantee of minimal impact of noisy labels for ground-truth labels. Empirical results on 20 imbalanced datasets with or without label noises demonstrate the outperformed performance of ANLI-FC in contrast to the comparative methods.
Keywords:
Antinoisy-labeling
area under the receiver operating characteristic (ROC) curve (AUC) maximization learning
fuzzy classifier with weighted fuzzy rules
generalization
imbalanced data classification
Journal
IF:
11.9
Papers:
5.1K
Citations:
2.9W

