Discriminative-Generative Positive and Unlabeled Learning.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Positive and Unlabeled (PU) learning trains classifiers using only positive and unlabeled data.
• Existing discriminative PU methods struggle with limited negative labels, hindering performance.

Purpose of the Study:

• To improve PU learning by integrating generative operations with discriminative methods.
• To address the challenge of insufficient supervisory information in PU learning.

Main Methods:

• Proposes a novel algorithm, Discriminative-Generative Positive and Unlabeled Learning (DGPU).
• Employs a tailored diffusion model for generating high-quality positive and negative examples.
• Iteratively refines a classifier through data generation and discriminative annotation stages.

Main Results:

• DGPU significantly outperforms existing PU methods on benchmark datasets.
• Achieves performance comparable to fully supervised methods.
• Improves test accuracy by 3.89% on CIFAR-10 and 2.56% on CelebA.

Conclusions:

• The integration of diffusion models into PU learning offers a synergistic benefit between generative and discriminative models.
• DGPU represents a significant advancement in PU learning, offering near fully supervised performance.