Robust AUC optimization under the supervision of clean data.

Area of Science:

• Machine Learning
• Data Mining

Background:

• Traditional area under the ROC curve (AUC) optimization requires large clean datasets, which are rare in real-world scenarios.
• Existing robust AUC optimization methods often neglect the utility of available clean data when dealing with noisy samples.

Purpose of the Study:

• To propose a novel framework for AUC optimization that effectively leverages both clean and noisy data.
• To enhance the robustness of AUC optimization in the presence of massive noisy samples.

Main Methods:

• A new framework for AUC optimization using self-paced learning (SPL) to guide noisy dataset processing with clean samples.
• Introduction of a consistency regularization term to mitigate the impact of data augmentation on SPL.
• Development of an efficient algorithm utilizing stochastic gradient methods for faster training by alternating updates of sample weights and model parameters.

Main Results:

• The proposed optimization method is theoretically proven to converge to a stationary point.
• The robust AUC optimization (RAUCO) algorithm demonstrates superior robustness compared to existing methods in comprehensive experiments.

Conclusions:

• The developed RAUCO algorithm offers a robust solution for AUC optimization with noisy datasets.
• The framework effectively utilizes clean data to improve the processing of noisy data, outperforming traditional and existing robust methods.