Robust AUC Optimization under the Supervision of Clean Data. (arXiv:2211.11751v1 [cs.LG])

AUC (area under the ROC curve) optimization algorithms have drawn much
attention due to the incredible adaptability for seriously imbalanced data.
Real-world datasets usually contain extensive noisy samples that seriously
hinder the model performance, but a limited number of clean samples can be
obtained easily. Although some AUC optimization studies make an effort to
dispose of noisy samples, they do not utilize such clean samples well. In this
paper, we propose a robust AUC optimization algorithm (RAUCO) with good use of
available clean samples. Expressly, our RAUCO algorithm can exclude noisy
samples from the training by employing the technology of self-paced learning
(SPL) under the supervision of clean samples. Moreover, considering the impact
of the data enhancement technology on SPL, we innovatively introduce the
consistency regularization term to SPL. Theoretical results on the convergence
of our RAUCO algorithm are provided under mild assumptions. Comprehensive
experiments demonstrate that our RAUCO algorithm holds better robustness than
existing algorithms.