cPU: Consistent Risk Estimator for Positive-Unlabeled Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Cost-sensitive methods are classical strategies for positive-unlabeled (PU) learning, aiming to minimize overall risk.
• Exploiting class priors is common in PU learning but effectively mining supervision information remains challenging.

Purpose of the Study:

• To derive a novel supervision formulation for PU learning based on risk minimization and class priors.
• To develop a consistent risk estimator that exploits the fixed distribution of samples in unlabeled data.

Main Methods:

• Derived a novel supervision formulation from a risk perspective, utilizing the convergence of risk distribution ratios in unlabeled samples.
• Constructed a consistent risk estimator to enforce consistency between negative and positive expected risks of respective samples.
• Introduced Mixup regularization to mitigate confirmation bias and entropy minimization to enhance sample separability.

Main Results:

• The proposed method demonstrates advantages over several baseline methods on four benchmark datasets.
• The novel formulation and risk estimator effectively mine underlying supervision information using class priors.
• Mixup regularization and entropy minimization further improved classification performance.

Conclusions:

• The developed PU learning approach effectively utilizes class priors for improved risk minimization and classification.
• The novel risk perspective and techniques like Mixup regularization offer a promising direction for PU learning.
• The method provides a consistent risk estimator and enhances model robustness and sample separability.