ComCo: Complementary supervised contrastive learning for complementary label learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Complementary label learning (CLL) reduces data labeling costs by using labels indicating what a sample is NOT.
• Existing CLL methods often neglect the rich semantic information within complementary labels.
• CLL remains a challenging problem despite its potential benefits.

Purpose of the Study:

• To introduce ComCo, a novel method that enhances complementary label learning (CLL) by incorporating contrastive learning.
• To address the limitations of previous CLL approaches by effectively utilizing semantic information from complementary labels.

Main Methods:

• ComCo employs a contrastive learning framework tailored for CLL.
• Key strategies include a positive selection mechanism for reliable positive samples.
• A negative selection strategy effectively utilizes complementary label information to construct informative negative sets.

Main Results:

• ComCo demonstrates significantly improved representation learning capabilities.
• The proposed method outperforms baseline models and current state-of-the-art approaches in CLL.
• Empirical results show performance gains of up to 14.61% in complementary label learning tasks.

Conclusions:

• ComCo offers a powerful new approach to complementary label learning by integrating contrastive learning.
• The method's ability to leverage semantic information in complementary labels leads to superior performance.
• ComCo represents a significant advancement in the field of machine learning for cost-effective data labeling.