Multi-Label Learning With Multiple Complementary Labels.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Multi-label classification assigns multiple labels to an instance.
• Complementary Label Learning (CLL) uses irrelevant labels to simplify annotation.
• Existing methods handle only single complementary labels per instance.

Purpose of the Study:

• Introduce a novel paradigm: Multi-label Learning with Multiple Complementary Labels (ML-MCL).
• Address limitations of existing methods in handling multiple complementary labels.
• Develop a robust and stable learning framework for ML-MCL tasks.

Main Methods:

• Analyze the generation process of multiple complementary labels.
• Construct the relationship between relevant and complementary labels.
• Derive a risk-consistent estimator with theoretical error bounds (O(1/sqrt(n))).
• Incorporate a confidence truncation loss to stabilize optimization.

Main Results:

• The proposed estimator achieves a convergence rate of O(1/sqrt(n)).
• Confidence truncation loss effectively mitigates unbounded gradients and stabilizes optimization.
• Experimental results demonstrate improved learning stability and performance.

Conclusions:

• ML-MCL is a practical extension of traditional complementary label learning.
• The proposed method effectively handles multiple complementary labels simultaneously.
• The approach offers enhanced stability and performance for multi-label classification.