Partial Label Learning via Gaussian Processes.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Probabilistic Modeling

Background:

• Partial label learning (PL) is a weakly supervised framework for data where each sample has a set of candidate labels.
• Precisely-labeled data is costly, making PL valuable for real-world applications.
• Existing PL algorithms face challenges due to inherent data ambiguity.

Purpose of the Study:

• To propose a novel probabilistic kernel algorithm for partial label learning.
• To address the ambiguity in training data within the PL framework.
• To enhance the accuracy of partial label learning models.

Main Methods:

• Employing a Gaussian process model with a prior on latent functions for each class.
• Defining a new likelihood function to disambiguate labeling information.
• Utilizing an aggregate function to approximate likelihood, resulting in a tighter max-loss equivalent and a differentiable convex objective function.

Main Results:

• The proposed Gaussian process-based algorithm demonstrates improved accuracy over existing state-of-the-art PL methods.
• Experiments conducted on six UCI datasets and three real-world PL problems validate the algorithm's effectiveness.
• The method successfully handles the ambiguity inherent in partial label learning.

Conclusions:

• The novel probabilistic kernel algorithm based on Gaussian processes offers a robust solution for partial label learning.
• The approach effectively disambiguates labels and achieves higher accuracy compared to current methods.
• This work contributes a significant advancement to the field of weakly supervised machine learning.