Learning With Proper Partial Labels.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial-label learning (PLL) uses sets of candidate labels instead of single true labels for training data.
• Existing PLL methods often rely on strong distributional assumptions for label generation models.
• Violating these assumptions can compromise the theoretical performance guarantees of current PLL algorithms.

Purpose of the Study:

• To introduce the concept of properness in partial labels for a more robust PLL framework.
• To develop a unified approach that encompasses various existing PLL settings.
• To establish theoretical guarantees for risk estimation in PLL.

Main Methods:

• Proposed the notion of properness on partial labels.
• Derived a unified unbiased estimator for classification risk.
• Proved the risk consistency and established an estimation error bound for the proposed estimator.

Main Results:

• The proper partial-label learning framework requires weaker distributional assumptions compared to prior methods.
• The proposed unified estimator is risk consistent.
• Experimental validation demonstrated the effectiveness of the developed algorithm.

Conclusions:

• The proper partial-label learning framework offers a theoretically sound and practically effective approach to weakly supervised learning.
• The unified estimator provides reliable risk estimation under weaker assumptions.
• This work advances the field of partial-label learning by providing a more general and robust methodology.