Learning From M-Tuple One-vs-All Confidence Comparison Data

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Partial-label learning (PLL) addresses classification with ambiguous supervision, where instances have candidate label sets containing the true label.
• Real-world annotation often provides richer structure than simple sets, including ranked candidates and potential omissions of the true label.

Purpose of the Study:

• To propose a novel weak-supervision paradigm, preferred-label PLL in a one-vs-all view (PLL-OVA), to handle more realistic annotation scenarios.
• To model preferred-label generation and incorporate missing-true-label situations for improved robustness.

Main Methods:

• Developed a flexible noisy-channel formulation for preferred-label generation, accommodating symmetric and asymmetric mis-selection.
• Derived principled empirical risk minimization (ERM) procedures via risk rewriting and established identifiability requirements.
• Introduced risk-correction functions (e.g., ReLU/ABS) to enhance optimization stability and robustness.

Main Results:

• PLL-OVA consistently outperforms standard PLL baselines across benchmark datasets.
• The proposed method shows significant improvements, particularly with high candidate ambiguity or non-uniform mis-selection patterns.
• Empirical performance is enhanced by the introduced risk-correction functions.

Conclusions:

• PLL-OVA offers a more practical and effective framework for partial-label learning by accounting for preferred labels and missing true labels.
• The method demonstrates robustness and superior performance in challenging, realistic noisy annotation settings.
• The study validates the effectiveness and practicality of the PLL-OVA framework for weak-supervised classification.