Large Margin Partial Label Machine.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial label learning (PLL) is a multi-class weakly supervised learning challenge with ambiguous labels.
• Large margin (LM)-based algorithms offer powerful disambiguation but often neglect candidate labels or use inaccurate capacity estimations, hindering generalization.

Purpose of the Study:

• To address the limitations of existing LM-based algorithms in PLL.
• To propose a novel LM-based algorithm that effectively handles label ambiguity without auxiliary estimations.

Main Methods:

• Developed LM-PLANE, extending multi-class support vector machines (SVM) to PLL.
• Incorporated optimistic superset loss to consider all potential candidate labels.
• Implemented an efficient cutting plane (CP) method for dual-space training.

Main Results:

• LM-PLANE demonstrated superior performance compared to existing LM-based and representative PLL algorithms.
• The proposed CP method showed effectiveness and convergence in theoretical analysis.
• Extensive experiments confirmed the algorithm's superiority across various PLL tasks.

Conclusions:

• LM-PLANE effectively disambiguates labels in PLL by considering all candidate labels.
• The algorithm achieves improved generalization performance over existing methods.
• The efficient CP training method provides a robust approach for LM-based PLL.