AUC-Oriented Domain Adaptation: From Theory to Algorithm.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• The Area Under the ROC curve (AUC) is vital for machine learning, especially in disease prediction and fraud detection with imbalanced datasets.
• Existing AUC methods often fail when training and test data distributions differ (domain shift).
• Addressing domain shift in AUC-oriented learning (AUCUDA) is an open challenge.

Purpose of the Study:

• To propose a theoretical framework and practical solution for AUC-oriented Unsupervised Domain Adaptation (AUCUDA).
• To develop a method that overcomes the limitations of standard AUC learning under domain shift.

Main Methods:

• Constructed a generalization bound using a novel distributional discrepancy for AUC.
• Developed a theoretical result addressing the interdependency of AUC risk terms.
• Proposed a pseudo-labeling strategy integrated into an end-to-end training framework to overcome annotation requirements.

Main Results:

• The proposed generalization bound offers sharper insights into AUC risk under domain shift.
• The end-to-end framework effectively handles domain shift without target domain annotations.
• Empirical studies on five real-world datasets demonstrate the framework's efficacy.

Conclusions:

• The developed AUCUDA framework provides a robust solution for machine learning tasks facing domain shift.
• This work lays the foundation for AUC-oriented domain adaptation research.
• The proposed methods show significant improvements in AUC performance across diverse datasets.