Biased Complementary-Label Learning Without True Labels.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Complementary-label learning (CLL) relies on transition matrices to correct label noise.
• Existing methods often assume identical transition probabilities, leading to bias.
• Biased complementary labels (CLs) present challenges due to corruption and unevenness.

Purpose of the Study:

• To develop a statistically consistent classifier for biased complementary-label learning.
• To propose a novel method for learning transition matrices without relying on anchor points or their posteriors.
• To address the limitations of existing methods in handling biased CLs.

Main Methods:

• Introduced a divided-T estimator to learn transition matrices from biased CLs.
• Utilized semantic clustering to mitigate the adverse effects of CLs.
• Factorized the transition matrix into two easily estimable matrices using learned semantic clusters as an intermediate class.

Main Results:

• The divided-T estimator effectively learns transition matrices under a mild assumption.
• Theoretical analyses and empirical results validate the estimator's effectiveness.
• Outperformed state-of-the-art methods on benchmark datasets.

Conclusions:

• The divided-T estimator provides a robust solution for learning transition matrices in biased CLL.
• Semantic clustering integration enhances the handling of noisy and uneven CLs.
• The proposed method offers a significant advancement over existing approaches in CLL.