Weakly Supervised AUC Optimization: A Unified Partial AUC Approach.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Real-world machine learning often faces imperfect data, known as weak supervision, including noisy, incomplete, or inexact labels.
• Traditional methods struggle with these data imperfections, hindering reliable performance evaluation and optimization.

Purpose of the Study:

• To develop a unified framework, WSAUC, for addressing Area Under the Curve (AUC) optimization under various weak supervision scenarios.
• To provide a robust and consistent approach for AUC maximization despite data contamination.

Main Methods:

• Framing AUC optimization in diverse weak supervision settings as minimizing AUC risk on contaminated datasets.
• Introducing reversed partial AUC (rpAUC) as a novel, robust training objective for AUC maximization.
• Developing the WSAUC framework to maximize empirical rpAUC for universal applicability.

Main Results:

• Demonstrated consistency between empirical risk minimization and true AUC under weak supervision.
• Empirical validation of WSAUC's effectiveness across noisy label, positive-unlabeled, multi-instance, and semi-supervised learning.
• Theoretical and experimental evidence supporting WSAUC's performance on multiple weakly supervised AUC optimization tasks.

Conclusions:

• WSAUC provides a universal and effective solution for AUC optimization in the presence of weak supervision.
• The proposed rpAUC metric offers a robust objective for training models with contaminated labels.
• The framework successfully integrates various weak supervision scenarios under a common AUC optimization approach.