Class-imbalanced complementary-label learning via weighted loss.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Complementary-label learning (CLL) is a weakly supervised classification method.
• Real-world datasets often exhibit class imbalance, degrading CLL performance.
• Existing CLL methods do not address class imbalance.

Purpose of the Study:

• To propose a novel problem setting for learning with class-imbalanced complementary labels in multi-class classification.
• To introduce Weighted Complementary-Label Learning (WCLL) to tackle this challenge.

Main Methods:

• Developed a novel Weighted Complementary-Label Learning (WCLL) approach.
• Modeled a weighted empirical risk minimization loss for imbalanced complementary labels.
• Derived an estimation error bound for theoretical guarantees.

Main Results:

• WCLL demonstrates significant improvements on benchmark and real-world datasets.
• The method effectively handles multi-class imbalanced scenarios.
• Achieved superior performance compared to state-of-the-art methods.

Conclusions:

• WCLL successfully addresses the challenge of class imbalance in complementary-label learning.
• The proposed method offers a robust solution for weakly supervised classification with imbalanced data.
• WCLL provides both practical performance gains and theoretical validation.