Partial label learning: Taxonomy, analysis and outlook.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial Label Learning (PLL) is an emerging weakly supervised machine learning framework.
• PLL addresses scenarios where training examples have a set of candidate labels, with only one true label.
• This framework has broad application prospects in various domains.

Purpose of the Study:

• To propose a novel taxonomy framework for Partial Label Learning (PLL).
• To categorize existing PLL methods into four main strategies: disambiguation, transformation, theory-oriented, and extensions.
• To provide a structured overview and facilitate future research in PLL.

Main Methods:

• Development of a new taxonomy for classifying Partial Label Learning (PLL) methods.
• Analysis and evaluation of existing PLL methods within the proposed categorical framework.
• Compilation and curation of synthetic and real-world PLL datasets with hyperlinks to source data.

Main Results:

• A comprehensive taxonomy for Partial Label Learning (PLL) is established, categorizing methods into four distinct strategies.
• Analysis reveals the strengths and weaknesses of methods across different categories.
• A curated collection of PLL datasets is provided for reproducible research.

Conclusions:

• The proposed taxonomy offers a structured approach to understanding and advancing Partial Label Learning (PLL).
• The framework facilitates the identification of research gaps and future directions in the field.
• The curated datasets will aid researchers in developing and evaluating new PLL algorithms.