Adaptive Graph Guided Disambiguation for Partial Label Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Partial label learning (PLL) uses ambiguous supervision where examples have multiple candidate labels.
• Label disambiguation aims to identify the true label from a set of candidates.
• Existing feature-aware methods struggle with noisy data affecting feature space graph structures.

Purpose of the Study:

• To propose a novel partial label learning approach using adaptive graph guided disambiguation.
• To enhance the reliability of feature space graph structures in the presence of noise and outliers.
• To integrate human-in-the-loop active learning for manual label disambiguation.

Main Methods:

• Developed an adaptive graph guided disambiguation method for partial label learning.
• Employed alternating optimization to jointly perform adaptive graph construction, label disambiguation, and model induction.
• Incorporated a human-in-the-loop framework for active querying of ambiguous examples.

Main Results:

• The proposed method effectively reveals intrinsic manifold structures among training examples.
• Adaptive graph guided disambiguation demonstrates superior performance compared to existing approaches.
• Extensive experiments validate the effectiveness of the approach in partial label learning scenarios.

Conclusions:

• Adaptive graph guided disambiguation is a robust strategy for learning from partial labels, especially with noisy data.
• Joint optimization of graph construction, disambiguation, and induction improves learning efficiency.
• The human-in-the-loop integration offers practical advantages for real-world applications.