Variational Label Enhancement for Instance-Dependent Partial Label Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Partial label learning (PLL) is a type of weakly supervised learning where each instance has a set of candidate labels, with only one being correct.
• Existing PLL methods often assume random selection of incorrect labels, which is unrealistic as labels can be instance-dependent.
• Instance-dependent label noise poses a significant challenge in real-world machine learning applications.

Purpose of the Study:

• To address the instance-dependent partial label learning problem.
• To propose novel methods that account for the instance-dependent nature of candidate labels in PLL.
• To improve the accuracy of predictive models in scenarios with non-random label noise.

Main Methods:

• Proposed two methods: VALEN and MILEN, designed to handle instance-dependent PLL.
• Both methods utilize latent label distributions recovered through a label enhancement process.
• VALEN infers variational posterior density using an inference model and evidence lower bound; MILEN uses variational approximation to bound mutual information.

Main Results:

• The proposed VALEN and MILEN methods demonstrate effectiveness in handling instance-dependent partial label learning.
• Experiments on benchmark and real-world datasets validate the superior performance of the proposed approaches.
• The methods successfully recover latent label distributions, leading to improved model training.

Conclusions:

• The developed methods, VALEN and MILEN, offer effective solutions for the challenging instance-dependent partial label learning problem.
• Accounting for instance-dependent label distributions is crucial for improving PLL performance.
• The proposed label enhancement techniques provide a robust framework for weakly supervised learning with complex noise patterns.